野村 行弘

In this study, we investigated the application of distributed learning, including federated learning and cyclical weight transfer, in the development of computer-aided detection (CADe) software for (1) cerebral aneurysm detection in magnetic resonance (MR) angiography images and (2) brain metastasis detection in brain contrast-enhanced MR images. We used datasets collected from various institutions, scanner vendors, and magnetic field strengths for each target CADe software. We compared the performance of multiple strategies, including a centralized strategy, in which software development is conducted at a development institution after collecting de-identified data from multiple institutions. Our results showed that the performance of CADe software trained through distributed learning was equal to or better than that trained through the centralized strategy. However, the distributed learning strategies that achieved the highest performance depend on the target CADe software. Hence, distributed learning can become one of the strategies for CADe software development using data collected from multiple institutions.

Abstract Purpose Many large radiographic datasets of lung nodules are available, but the small and hard-to-detect nodules are rarely validated by computed tomography. Such difficult nodules are crucial for training nodule detection methods. This lack of difficult nodules for training can be addressed by artificial nodule synthesis algorithms, which can create artificially embedded nodules. This study aimed to develop and evaluate a novel cost function for training networks to detect such lesions. Embedding artificial lesions in healthy medical images is effective when positive cases are insufficient for network training. Although this approach provides both positive (lesion-embedded) images and the corresponding negative (lesion-free) images, no known methods effectively use these pairs for training. This paper presents a novel cost function for segmentation-based detection networks when positive–negative pairs are available. Methods Based on the classic U-Net, new terms were added to the original Dice loss for reducing false positives and the contrastive learning of diseased regions in the image pairs. The experimental network was trained and evaluated, respectively, on 131,072 fully synthesized pairs of images simulating lung cancer and real chest X-ray images from the Japanese Society of Radiological Technology dataset. Results The proposed method outperformed RetinaNet and a single-shot multibox detector. The sensitivities were 0.688 and 0.507 when the number of false positives per image was 0.2, respectively, with and without fine-tuning under the leave-one-case-out setting. Conclusion To our knowledge, this is the first study in which a method for detecting pulmonary nodules in chest X-ray images was evaluated on a real clinical dataset after being trained on fully synthesized images. The synthesized dataset is available at https://zenodo.org/records/10648433.

Abstract Purpose To assess the performance of GPT-4 Turbo with Vision (GPT-4TV), OpenAI’s latest multimodal large language model, by comparing its ability to process both text and image inputs with that of the text-only GPT-4 Turbo (GPT-4 T) in the context of the Japan Diagnostic Radiology Board Examination (JDRBE). Materials and methods The dataset comprised questions from JDRBE 2021 and 2023. A total of six board-certified diagnostic radiologists discussed the questions and provided ground-truth answers by consulting relevant literature as necessary. The following questions were excluded: those lacking associated images, those with no unanimous agreement on answers, and those including images rejected by the OpenAI application programming interface. The inputs for GPT-4TV included both text and images, whereas those for GPT-4 T were entirely text. Both models were deployed on the dataset, and their performance was compared using McNemar’s exact test. The radiological credibility of the responses was assessed by two diagnostic radiologists through the assignment of legitimacy scores on a five-point Likert scale. These scores were subsequently used to compare model performance using Wilcoxon's signed-rank test. Results The dataset comprised 139 questions. GPT-4TV correctly answered 62 questions (45%), whereas GPT-4 T correctly answered 57 questions (41%). A statistical analysis found no significant performance difference between the two models (P = 0.44). The GPT-4TV responses received significantly lower legitimacy scores from both radiologists than the GPT-4 T responses. Conclusion No significant enhancement in accuracy was observed when using GPT-4TV with image input compared with that of using text-only GPT-4 T for JDRBE questions.

Abstract To efficiently allocate medical resources at disaster sites, medical workers perform triage to prioritize medical treatments based on the severity of the wounded or sick. In such instances, evaluators often assess the severity status of the wounded or sick quickly, but their measurements are qualitative and rely on experience. Therefore, we developed a wearable device called Medic Hand in this study to extend the functionality of a medical worker’s hand so as to measure multiple biometric indicators simultaneously without increasing the number of medical devices to be carried. Medic Hand was developed to quantitatively and efficiently evaluate "perfusion" during triage. Speed is essential during triage at disaster sites, where time and effort are often spared to attach medical devices to patients, so the use of Medic Hand as a biometric measurement device is more efficient for collecting biometric information. For Medic Hand to be handy during disasters, it is essential to understand and improve upon factors that facilitate its public acceptance. To this end, this paper reports on the usability evaluation of Medic Hand through a questionnaire survey of nonmedical workers.

Local differential privacy algorithms combined with deep generative models can enhance secure medical image sharing among researchers in the public domain without central administrators; however, these images were limited to the generation of low-resolution images, which are very insufficient for diagnosis by medical doctors. To enhance the performance of deep generative models so that they can generate high-resolution medical images, we propose a large-scale diffusion model that can, for the first time, unconditionally generate high-resolution (256×256×256) volumetric medical images (head magnetic resonance images). This diffusion model has 19 billion parameters, but to make it easy to train it, we temporally divided the model into 200 submodels, each of which has 95 million parameters. Moreover, on the basis of this new diffusion model, we propose another formulation of image anonymization with which the processed images can satisfy provable Gaussian local differential privacy and with which we can generate images semantically different from the original image but belonging to the same class. We believe that the formulation of this new diffusion model and the implementation of local differential privacy algorithms combined with the diffusion models can contribute to the secure sharing of practical images upstream of data processing.

Abstract Purpose The quality and bias of annotations by annotators (e.g., radiologists) affect the performance changes in computer-aided detection (CAD) software using machine learning. We hypothesized that the difference in the years of experience in image interpretation among radiologists contributes to annotation variability. In this study, we focused on how the performance of CAD software changes with retraining by incorporating cases annotated by radiologists with varying experience. Methods We used two types of CAD software for lung nodule detection in chest computed tomography images and cerebral aneurysm detection in magnetic resonance angiography images. Twelve radiologists with different years of experience independently annotated the lesions, and the performance changes were investigated by repeating the retraining of the CAD software twice, with the addition of cases annotated by each radiologist. Additionally, we investigated the effects of retraining using integrated annotations from multiple radiologists. Results The performance of the CAD software after retraining differed among annotating radiologists. In some cases, the performance was degraded compared to that of the initial software. Retraining using integrated annotations showed different performance trends depending on the target CAD software, notably in cerebral aneurysm detection, where the performance decreased compared to using annotations from a single radiologist. Conclusions Although the performance of the CAD software after retraining varied among the annotating radiologists, no direct correlation with their experience was found. The performance trends differed according to the type of CAD software used when integrated annotations from multiple radiologists were used.

Abstract Background Previous research applying large language models (LLMs) to medicine was focused on text-based information. Recently, multimodal variants of LLMs acquired the capability of recognizing images. Objective To evaluate the capability of GPT-4V, a recent multimodal LLM developed by OpenAI, in recognizing images in the medical field by testing its capability to answer questions in the 117th Japanese National Medical Licensing Examination. Methods We focused on 108 questions that had one or more images as part of a question and presented GPT-4V with the same questions under two conditions: 1) with both the question text and associated image(s), and 2) with the question text only. We then compared the difference in accuracy between the two conditions using the exact McNemar’s test. Results Among the 108 questions with images, GPT-4V’s accuracy was 68% when presented with images and 72% when presented without images (P= .36). Conclusions The additional information from the images did not significantly improve the performance of GPT-4V in the Japanese Medical Licensing Examination.

An educational augmented reality auscultation system (EARS) is proposed to enhance the reality of auscultation training using a simulated patient. The conventional EARS cannot accurately reproduce breath sounds according to the breathing of a simulated patient because the system instructs the breathing rhythm. In this study, we propose breath measurement methods that can be integrated into the chest piece of a stethoscope. We investigate methods using the thoracic variations and frequency characteristics of breath sounds. An accelerometer, a magnetic sensor, a gyro sensor, a pressure sensor, and a microphone were selected as the sensors. For measurement with the magnetic sensor, we proposed a method by detecting the breathing waveform in terms of changes in the magnetic field accompanying the surface deformation of the stethoscope based on thoracic variations using a magnet. During breath sound measurement, the frequency spectra of the breath sounds acquired by the built-in microphone were calculated. The breathing waveforms were obtained from the difference in characteristics between the breath sounds during exhalation and inhalation. The result showed the average value of the correlation coefficient with the reference value reached 0.45, indicating the effectiveness of this method as a breath measurement method. And the evaluations suggest more accurate breathing waveforms can be obtained by selecting the measurement method according to breathing method and measurement point.

The purpose of the study was to develop a liver nodule diagnostic method that accurately localizes and classifies focal liver lesions and identifies the specific liver segments in which they reside by integrating a liver segment division algorithm using a four-dimensional (4D) fully convolutional residual network (FC-ResNet) with a localization and classification model. We retrospectively collected data and divided 106 gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid-enhanced magnetic resonance examinations into Case-sets 1, 2, and 3. A liver segment division algorithm was developed using a 4D FC-ResNet and trained with semi-automatically created silver-standard annotations; performance was evaluated using manually created gold-standard annotations by calculating the Dice scores for each liver segment. The performance of the liver nodule diagnostic method was assessed by comparing the results with those of the original radiology reports. The mean Dice score between the output of the liver segment division model and the gold standard was 0.643 for Case-set 2 (normal liver contours) and 0.534 for Case-set 1 (deformed liver contours). Among the 64 lesions in Case-set 3, the diagnostic method localized 37 lesions, classified 33 lesions, and identified the liver segments for 30 lesions. A total of 28 lesions were true positives, matching the original radiology reports. The liver nodule diagnostic method, which integrates a liver segment division algorithm with a lesion localization and classification model, exhibits great potential for localizing and classifying focal liver lesions and identifying the liver segments in which they reside. Further improvements and validation using larger sample sizes will enhance its performance and clinical applicability.

PURPOSE: Standardized uptake values (SUVs) derived from 18F-fluoro-2-deoxy-D-glucose positron emission tomography/computed tomography are a crucial parameter for identifying tumors or abnormalities in an organ. Moreover, exploring ways to improve the identification of tumors or abnormalities using a statistical measurement tool is important in clinical research. Therefore, we developed a fully automatic method to create a personally normalized Z-score map of the liver SUV. METHODS: The normalized Z-score map for each patient was created using the SUV mean and standard deviation estimated from blood-test-derived variables, such as alanine aminotransferase and aspartate aminotransferase, as well as other demographic information. This was performed using the least absolute shrinkage and selection operator (LASSO)-based estimation formula. We also used receiver operating characteristic (ROC) to analyze the results of people with and without hepatic tumors and compared them to the ROC curve of normal SUV. RESULTS: A total of 7757 people were selected for this study. Of these, 7744 were healthy, while 13 had abnormalities. The area under the ROC curve results indicated that the anomaly detection approach (0.91) outperformed only the maximum SUV (0.89). To build the LASSO regression, sets of covariates, including sex, weight, body mass index, blood glucose level, triglyceride, total cholesterol, γ-glutamyl transpeptidase, total protein, creatinine, insulin, albumin, and cholinesterase, were used to determine the SUV mean, whereas weight was used to determine the SUV standard deviation. CONCLUSION: The Z-score normalizes the mean and standard deviation. It is effective in ROC curve analysis and increases the clarity of the abnormality. This normalization is a key technique for effective measurement of maximum glucose consumption by tumors in the liver.

To generate synthetic medical data incorporating image-tabular hybrid data by merging an image encoding/decoding model with a table-compatible generative model and assess their utility. We used 1342 cases from the Stony Brook University Covid-19-positive cases, comprising chest X-ray radiographs (CXRs) and tabular clinical data as a private dataset (pDS). We generated a synthetic dataset (sDS) through the following steps: (I) dimensionally reducing CXRs in the pDS using a pretrained encoder of the auto-encoding generative adversarial networks (αGAN) and integrating them with the correspondent tabular clinical data; (II) training the conditional tabular GAN (CTGAN) on this combined data to generate synthetic records, encompassing encoded image features and clinical data; and (III) reconstructing synthetic images from these encoded image features in the sDS using a pretrained decoder of the αGAN. The utility of sDS was assessed by the performance of the prediction models for patient outcomes (deceased or discharged). For the pDS test set, the area under the receiver operating characteristic (AUC) curve was calculated to compare the performance of prediction models trained separately with pDS, sDS, or a combination of both. We created an sDS comprising CXRs with a resolution of 256 × 256 pixels and tabular data containing 13 variables. The AUC for the outcome was 0.83 when the model was trained with the pDS, 0.74 with the sDS, and 0.87 when combining pDS and sDS for training. Our method is effective for generating synthetic records consisting of both images and tabular clinical data.

This study aimed to explore the relationship between thyroid-stimulating hormone (TSH) elevation and the baseline computed tomography (CT) density and volume of the thyroid. We examined 86 cases with new-onset hypothyroidism (TSH > 4.5 IU/mL) and 1071 controls from a medical check-up database over 5 years. A deep learning-based thyroid segmentation method was used to assess CT density and volume. Statistical tests and logistic regression were employed to determine differences and odds ratios. Initially, the case group showed a higher CT density (89.8 vs. 81.7 Hounsfield units (HUs)) and smaller volume (13.0 vs. 15.3 mL) than those in the control group. For every +10 HU in CT density and −3 mL in volume, the odds of developing hypothyroidism increased by 1.40 and 1.35, respectively. Over the course of the study, the case group showed a notable CT density reduction (median: −8.9 HU), whereas the control group had a minor decrease (−2.9 HU). Thyroid volume remained relatively stable for both groups. Higher CT density and smaller thyroid volume at baseline are correlated with future TSH elevation. Over time, there was a substantial and minor decrease in CT density in the case and control groups, respectively. Thyroid volumes remained consistent in both cohorts.

Lower extremity lymphedema (LEL) is a common complication after gynecological cancer treatment, which significantly reduces the quality of life. While early diagnosis and intervention can prevent severe complications, there is currently no consensus on the optimal screening strategy for postoperative LEL. In this study, we developed a computer-aided diagnosis (CAD) software for LEL screening in pelvic computed tomography (CT) images using deep learning. A total of 431 pelvic CT scans from 154 gynecological cancer patients were used for this study. We employed ResNet-18, ResNet-34, and ResNet-50 models as the convolutional neural network (CNN) architecture. The input image for the CNN model used a single CT image at the greater trochanter level. Fat-enhanced images were created and used as input to improve classification performance. Receiver operating characteristic analysis was used to evaluate our method. The ResNet-34 model with fat-enhanced images achieved the highest area under the curve of 0.967 and an accuracy of 92.9%. Our CAD software enables LEL diagnosis from a single CT image, demonstrating the feasibility of LEL screening only on CT images after gynecologic cancer treatment. To increase the usefulness of our CAD software, we plan to validate it using external datasets.

Diagnostic radiologists need artificial intelligence (AI) for medical imaging, but access to medical images required for training in AI has become increasingly restrictive. To release and use medical images, we need an algorithm that can simultaneously protect privacy and preserve pathologies in medical images. To address this, we introduce DP-GLOW, a hybrid that combines the local differential privacy (LDP) algorithm with GLOW, one of the flow-based deep generative models. By applying a GLOW model, we disentangle the pixelwise correlation of images, which makes it difficult to protect privacy with straightforward LDP algorithms for images. Specifically, we map images to the latent vector of the GLOW model, where each element follows an independent normal distribution. We then apply the Laplace mechanism to this latent vector to achieve ϵ-LDP, which is one of the LDP algorithms. Moreover, we applied DP-GLOW to chest X-ray images to generate LDP images while preserving pathologies. The ϵ-LDP-processed chest X-ray images obtained with DP-GLOW indicate that we have obtained a powerful tool for releasing and using medical images for training AI.

Importance Characterizing longitudinal patterns of regional brain volume changes in a population with normal cognition at the individual level could improve understanding of the brain aging process and may aid in the prevention of age-related neurodegenerative diseases. Objective To investigate age-related trajectories of the volumes and volume change rates of brain structures in participants without dementia. Design, Setting, and Participants This cohort study was conducted from November 1, 2006, to April 30, 2021, at a single academic health-checkup center among 653 individuals who participated in a health screening program with more than 10 years of serial visits. Exposure Serial magnetic resonance imaging, Mini-Mental State Examination, health checkup. Main Outcomes and Measures Volumes and volume change rates across brain tissue types and regions. Results The study sample included 653 healthy control individuals (mean [SD] age at baseline, 55.1 [9.3] years; median age, 55 years [IQR, 47-62 years]; 447 men [69%]), who were followed up annually for up to 15 years (mean [SD], 11.5 [1.8] years; mean [SD] number of scans, 12.1 [1.9]; total visits, 7915). Each brain structure showed characteristic age-dependent volume and atrophy change rates. In particular, the cortical gray matter showed a consistent pattern of volume loss in each brain lobe with aging. The white matter showed an age-related decrease in volume and an accelerated atrophy rate (regression coefficient, −0.016 [95% CI, −0.012 to –0.011]; P < .001). An accelerated age-related volume increase in the cerebrospinal fluid–filled spaces, particularly in the inferior lateral ventricle and the Sylvian fissure, was also observed (ventricle regression coefficient, 0.042 [95% CI, 0.037-0.047]; P < .001; sulcus regression coefficient, 0.021 [95% CI, 0.018-0.023]; P < .001). The temporal lobe atrophy rate accelerated from approximately 70 years of age, preceded by acceleration of atrophy in the hippocampus and amygdala. Conclusions and Relevance In this cohort study of adults without dementia, age-dependent brain structure volumes and volume change rates in various brain structures were characterized using serial magnetic resonance imaging scans. These findings clarified the normal distributions in the aging brain, which are essential for understanding the process of age-related neurodegenerative diseases.

INTRODUCTION: Capillary refill time (CRT) is an indicator of peripheral circulation and is recommended in the 2021 guidelines for treating and managing sepsis. STUDY OBJECTIVE: This study developed a portable device to realize objective CRT measurement. Assuming that peripheral blood flow obstruction by the artery occlusion test (AOT) or venous occlusion test (VOT) increases the CRT, the cut-off value for peripheral circulatory failure was studied by performing a comparative analysis with CRT with no occlusion test (No OT). METHODS: Fourteen (14) healthy adults (age: 20-26 years) participated in the study. For the vascular occlusion test, a sphygmomanometer was placed on the left upper arm of the participant in the supine position, and a pressure of 30mmHg higher than the systolic pressure was applied for AOT, a pressure of 60mmHg was applied for VOT, respectively, and no pressure was applied for No OT. The CRT was measured from the index finger of the participant's left hand. RESULTS: Experimental results revealed that CRT was significantly longer in the AOT and did not differ significantly in the VOT. The cut-off value for peripheral circulatory failure was found to be 2.88 seconds based on Youden's index by using receiver operating characteristic (ROC) analysis with AOT as positive and No OT as negative. CONCLUSION: Significant results were obtained in a previous study on the evaluation of septic shock patients when CRT > three seconds was considered abnormal, and the cut-off value for peripheral circulatory failure in the current study validated this.

The purpose of this study was to realize an automated volume measurement of abdominal adipose tissue from the entire abdominal cavity in Dixon magnetic resonance (MR) images using deep learning. Our algorithm involves a combination of extraction of the abdominal cavity and body trunk regions using deep learning and extraction of a fat region based on automatic thresholding. To evaluate the proposed method, we calculated the Dice coefficient (DC) between the extracted regions using deep learning and labeled images. We also compared the visceral adipose tissue (VAT) and subcutaneous adipose tissue volumes calculated by employing the proposed method with those calculated from computed tomography (CT) images scanned on the same day using the automatic calculation method previously developed by our group. We implemented our method as a plug-in in a web-based medical image processing platform. The DCs of the abdominal cavity and body trunk regions were 0.952 ± 0.014 and 0.995 ± 0.002, respectively. The VAT volume measured from MR images using the proposed method was almost equivalent to that measured from CT images. The time required for our plug-in to process the test set was 118.9 ± 28.0 s. Using our proposed method, the VAT volume measured from MR images can be an alternative to that measured from CT images.

OBJECTIVE: To determine a standardized cutoff value for abnormal 18F-fluorodeoxyglucose (FDG) accumulation in the thyroid gland. METHODS: Herein, 7013 FDG-positron emission tomography (PET)/computed tomography (CT) scans were included. An automatic thyroid segmentation method using two U-nets (2D- and 3D-U-net) was constructed; mean FDG standardized uptake value (SUV), CT value, and volume of the thyroid gland were obtained from each participant. The values were categorized by thyroid function into three groups based on serum thyroid stimulating hormone levels. Thyroid function and mean SUV with increments of 1 were analyzed, and risk for thyroid dysfunction was calculated. Thyroid dysfunction detection ability was examined using a machine learning method (Lightgbm) with age, sex, height, weight, CT value, volume, and mean SUV as explanatory variables. RESULTS: Mean SUV was significantly higher in females with hypothyroidism. Almost 98.9% of participants in the normal group had mean SUV <2 and 93.8% participants with mean SUV <2 had normal thyroid function. The hypothyroidism group had more cases with mean SUV ≥2. The relative risk of having abnormal thyroid function was 4.6 with mean SUV ≥2. The sensitivity and specificity for detecting thyroid dysfunction using Lightgbm were 14.5% and 99%, respectively. CONCLUSIONS: Mean SUV ≥2 was strongly associated with abnormal thyroid function in this large cohort, indicating that mean SUV with FDG-PET/CT can be used as a criterion for thyroid evaluation. Preliminarily, this study shows the potential utility of detecting thyroid dysfunction based on imaging findings.

Abstract Meteorological-tsunami-like (or meteotsunami-like) periodic oscillation was muographically detected with the Tokyo-Bay Seafloor Hyper-Kilometric Submarine Deep Detector (TS-HKMSDD) deployed in the underwater highway called the Trans-Tokyo Bay Expressway or Tokyo Bay Aqua-Line (TBAL). It was detected right after the arrival of the 2021 Typhoon-16 that passed through the region 400 km south of the bay. The measured oscillation period and decay time were respectively 3 h and 10 h. These measurements were found to be consistent with previous tide gauge measurements. Meteotsunamis are known to take place in bays and lakes, and the temporal and spatial characteristics of meteotsunamis are similar to seismic tsunamis. However, their generation and propagation mechanisms are not well understood. The current result indicates that a combination of muography and trans-bay or trans-lake underwater tunnels will offer an additional tool to measure meteotsunamis at locations where tide gauges are unavailable.

PURPOSE: To assess the clinical effectiveness of temporal subtraction computed tomography (TS CT) using deep learning to improve vertebral bone metastasis detection. METHOD: This retrospective study used TS CT comprising bony landmark detection, bone segmentation with a multi-atlas-based method, and spatial registration of two images by a log-domain diffeomorphic Demons algorithm. Paired current and past CT images of 50 patients without vertebral metastasis, recorded during June 2011-September 2016, were included as training data. A deep learning-based method estimated registration errors and suppressed false positives. Thereafter, paired CT images of 40 cancer patients with newly developed vertebral metastases and 40 control patients without vertebral metastases were evaluated. Six board-certified radiologists and five radiology residents independently interpreted 80 paired CT images with and without TS CT. RESULTS: Records of 40 patients in the metastasis group (median age: 64.5 years; 20 males) and 40 patients in the control group (median age: 64.0 years; 20 males) were evaluated. With TS CT, the overall figure of merit (FOM) of the board-certified radiologist and resident groups improved from 0.848 to 0.876 (p = 0.01) and from 0.752 to 0.799 (p = 0.02), respectively. The sub-analysis focusing on attenuation changes in lesions revealed that the FOM of osteoblastic lesions significantly improved in both the board-certified radiologist and resident groups using TS CT. The sub-analysis focusing on lesion location showed that the FOM of the resident group significantly improved in the vertebral arch (p = 0.04). CONCLUSIONS: TS CT was effective in detecting bone metastasis by both board-certified radiologists and radiology residents.

Ultra-sparse-view computed tomography (CT) algorithms can reduce radiation exposure for patients, but these algorithms lack an explicit cycle consistency loss minimization and an explicit log-likelihood maximization in testing. Here, we propose X2CT-FLOW for the maximum a posteriori (MAP) reconstruction of a three-dimensional (3D) chest CT image from a single or a few two-dimensional (2D) projection images using a progressive flow-based deep generative model, especially for ultra-low-dose protocols. The MAP reconstruction can simultaneously optimize the cycle consistency loss and the log-likelihood. We applied X2CT-FLOW for the reconstruction of 3D chest CT images from biplanar projection images without noise contamination (assuming a standard-dose protocol) and with strong noise contamination (assuming an ultra-low-dose protocol). We simulated an ultra-low-dose protocol. With the standard-dose protocol, our images reconstructed from 2D projected images and 3D ground-truth CT images showed good agreement in terms of structural similarity (SSIM, 0.7675 on average), peak signal-to-noise ratio (PSNR, 25.89 dB on average), mean absolute error (MAE, 0.02364 on average), and normalized root mean square error (NRMSE, 0.05731 on average). Moreover, with the ultra-low-dose protocol, our images reconstructed from 2D projected images and the 3D ground-truth CT images also showed good agreement in terms of SSIM (0.7008 on average), PSNR (23.58 dB on average), MAE (0.02991 on average), and NRMSE (0.07349 on average).

PURPOSE: To develop an anomaly detection system in PET/CT with the tracer 18F-fluorodeoxyglucose (FDG) that requires only normal PET/CT images for training and can detect abnormal FDG uptake at any location in the chest region. MATERIALS AND METHODS: We trained our model based on a Bayesian deep learning framework using 1878 PET/CT scans with no abnormal findings. Our model learns the distribution of standard uptake values in these normal training images and detects out-of-normal uptake regions. We evaluated this model using 34 scans showing focal abnormal FDG uptake in the chest region. This evaluation dataset includes 28 pulmonary and 17 extrapulmonary abnormal FDG uptake foci. We performed per-voxel and per-slice receiver operating characteristic (ROC) analyses and per-lesion free-response receiver operating characteristic analysis. RESULTS: Our model showed an area under the ROC curve of 0.992 on discriminating abnormal voxels and 0.852 on abnormal slices. Our model detected 41 of 45 (91.1%) of the abnormal FDG uptake foci with 12.8 false positives per scan (FPs/scan), which include 26 of 28 pulmonary and 15 of 17 extrapulmonary abnormalities. The sensitivity at 3.0 FPs/scan was 82.2% (37/45). CONCLUSION: Our model trained only with normal PET/CT images successfully detected both pulmonary and extrapulmonary abnormal FDG uptake in the chest region.

Medical artificial intelligence (AI) systems need to learn to recognize synonyms or paraphrases describing the same anatomy, disease, treatment, etc. to better understand real-world clinical documents. Existing linguistic resources focus on variants at the word or sentence level. To handle linguistic variations on a broader scale, we proposed the Medical Text Radiology Report section Japanese version (MedTxt-RR-JA), the first clinical comparable corpus. MedTxt-RR-JA was built by recruiting nine radiologists to diagnose the same 15 lung cancer cases in Radiopaedia, an open-access radiological repository. The 135 radiology reports in MedTxt-RR-JA were shown to contain word-, sentence- and document-level variations maintaining similarity of contents. MedTxt-RR-JA is also the first publicly available Japanese radiology report corpus that would help to overcome poor data availability for Japanese medical AI systems. Moreover, our methodology can be applied widely to building clinical corpora without privacy concerns.

In the original version of this Article, Masaki Satoh was incorrectly listed as an author of the original Article, and has subsequently been removed. The original Article has been corrected.

PURPOSE: The three-dimensional (3D) voxel labeling of lesions requires significant radiologists' effort in the development of computer-aided detection software. To reduce the time required for the 3D voxel labeling, we aimed to develop a generalized semiautomatic segmentation method based on deep learning via a data augmentation-based domain generalization framework. In this study, we investigated whether a generalized semiautomatic segmentation model trained using two types of lesion can segment previously unseen types of lesion. METHODS: We targeted lung nodules in chest CT images, liver lesions in hepatobiliary-phase images of Gd-EOB-DTPA-enhanced MR imaging, and brain metastases in contrast-enhanced MR images. For each lesion, the 32 × 32 × 32 isotropic volume of interest (VOI) around the center of gravity of the lesion was extracted. The VOI was input into a 3D U-Net model to define the label of the lesion. For each type of target lesion, we compared five types of data augmentation and two types of input data. RESULTS: For all considered target lesions, the highest dice coefficients among the training patterns were obtained when using a combination of the existing data augmentation-based domain generalization framework and random monochrome inversion and when using the resized VOI as the input image. The dice coefficients were 0.639 ± 0.124 for the lung nodules, 0.660 ± 0.137 for the liver lesions, and 0.727 ± 0.115 for the brain metastases. CONCLUSIONS: Our generalized semiautomatic segmentation model could label unseen three types of lesion with different contrasts from the surroundings. In addition, the resized VOI as the input image enables the adaptation to the various sizes of lesions even when the size distribution differed between the training set and the test set.

Tidal measurements are of great significance since they may provide us with essential data to apply towards protection of coastal communities and sea traffic. Currently, tide gauge stations and laser altimetry are commonly used for these measurements. On the other hand, muography sensors can be located underneath the seafloor inside an undersea tunnel where electric and telecommunication infrastructures are more readily available. In this work, the world's first under-seafloor particle detector array called the Tokyo-bay Seafloor Hyper-Kilometric Submarine Deep Detector (TS-HKMSDD) was deployed underneath the Tokyo-Bay seafloor for conducting submarine muography. The resultant 80-day consecutive time-sequential muographic data were converted to the tidal levels based on the parameters determined from the first-day astronomical tide height (ATH) data. The standard deviation between ATH and muographic results for the rest of a 79-day measurement period was 12.85 cm. We anticipate that if the length of the TS-HKMSDD is extended from 100 m to a full-scale as large as 9.6 km to provide continuous tidal information along the tunnel, this muography application will become an established standard, demonstrating its effectiveness as practical tide monitor for this heavy traffic waterway in Tokyo and in other important sea traffic areas worldwide.

BACKGROUND: It is essential for radiologists to communicate actionable findings to the referring clinicians reliably. Natural language processing (NLP) has been shown to help identify free-text radiology reports including actionable findings. However, the application of recent deep learning techniques to radiology reports, which can improve the detection performance, has not been thoroughly examined. Moreover, free-text that clinicians input in the ordering form (order information) has seldom been used to identify actionable reports. This study aims to evaluate the benefits of two new approaches: (1) bidirectional encoder representations from transformers (BERT), a recent deep learning architecture in NLP, and (2) using order information in addition to radiology reports. METHODS: We performed a binary classification to distinguish actionable reports (i.e., radiology reports tagged as actionable in actual radiological practice) from non-actionable ones (those without an actionable tag). 90,923 Japanese radiology reports in our hospital were used, of which 788 (0.87%) were actionable. We evaluated four methods, statistical machine learning with logistic regression (LR) and with gradient boosting decision tree (GBDT), and deep learning with a bidirectional long short-term memory (LSTM) model and a publicly available Japanese BERT model. Each method was used with two different inputs, radiology reports alone and pairs of order information and radiology reports. Thus, eight experiments were conducted to examine the performance. RESULTS: Without order information, BERT achieved the highest area under the precision-recall curve (AUPRC) of 0.5138, which showed a statistically significant improvement over LR, GBDT, and LSTM, and the highest area under the receiver operating characteristic curve (AUROC) of 0.9516. Simply coupling the order information with the radiology reports slightly increased the AUPRC of BERT but did not lead to a statistically significant improvement. This may be due to the complexity of clinical decisions made by radiologists. CONCLUSIONS: BERT was assumed to be useful to detect actionable reports. More sophisticated methods are required to use order information effectively.

PURPOSE: Gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA)-enhanced magnetic resonance imaging (MRI) has high diagnostic accuracy in the detection of liver lesions. There is a demand for computer-aided detection/diagnosis software for Gd-EOB-DTPA-enhanced MRI. We propose a deep learning-based method using one three-dimensional fully convolutional residual network (3D FC-ResNet) for liver segmentation and another 3D FC-ResNet for simultaneous detection and classification of a focal liver lesion in Gd-EOB-DTPA-enhanced MRI. METHODS: We prepared a five-phase (unenhanced, arterial, portal venous, equilibrium, and hepatobiliary phases) series as the input image sets and labeled focal liver lesion (hepatocellular carcinoma, metastasis, hemangiomas, cysts, and scars) images as the output image sets. We used 100 cases to train our model, 42 cases to determine the hyperparameters of our model, and 42 cases to evaluate our model. We evaluated our model by free-response receiver operating characteristic curve analysis and using a confusion matrix. RESULTS: Our model simultaneously detected and classified focal liver lesions. In the test cases, the detection accuracy for whole focal liver lesions had a true-positive ratio of 0.6 at an average of 25 false positives per case. The classification accuracy was 0.790. CONCLUSION: We proposed the simultaneous detection and classification of a focal liver lesion in Gd-EOB-DTPA-enhanced MRI using multichannel 3D FC-ResNet. Our results indicated simultaneous detection and classification are possible using a single network. It is necessary to further improve detection sensitivity to help radiologists.

PURPOSE: Radiologists interpret many medical images and clinical practice demands timely interpretation, resulting in a heavy workload. To reduce the workload, here we formulate and validate a method that can handle different types of medical image and can detect virtually all types of lesion in a medical image. For the first time, we show that two flow-based deep generative (FDG) models can predict the logarithm posterior probability in a semi-supervised approach. METHODS: We adopt two FDG models in conjunction with Bayes' theorem to predict the logarithm posterior probability that a medical image is normal. We trained one of the FDG models with normal images and the other FDG model with normal and non-normal images. RESULTS: We validated the method using two types of medical image: chest X-ray images (CXRs) and brain computed tomography images (BCTs). The area under the receiver operating characteristic curve for pneumonia-like opacities in CXRs was 0.839 on average, and for infarction in BCTs was 0.904. CONCLUSION: We formulated a method of predicting the logarithm posterior probability using two FDG models. We validated that the method can detect abnormal findings in CXRs and BCTs with both an acceptable performance for testing and a comparatively light workload for training.

PURPOSE: The clinical usefulness of computer-aided detection of cerebral aneurysms has been investigated using different methods to present lesion candidates, but suboptimal methods may have limited its usefulness. We compared three presentation methods to determine which can benefit radiologists the most by enabling them to detect more aneurysms. MATERIALS AND METHODS: We conducted a multireader multicase observer performance study involving six radiologists and using 470 lesion candidates output by a computer-aided detection program, and compared the following three different presentation methods using the receiver operating characteristic analysis: (1) a lesion candidate is encircled on axial slices, (2) a lesion candidate is overlaid on a volume-rendered image, and (3) combination of (1) and (2). The response time was also compared. RESULTS: As compared with axial slices, radiologists showed significantly better detection performance when presented with volume-rendered images. There was no significant difference in response time between the two methods. The combined method was associated with a significantly longer response time, but had no added merit in terms of diagnostic accuracy. CONCLUSION: Even with the aid of computer-aided detection, radiologists overlook many aneurysms if the presentation method is not optimal. Overlaying colored lesion candidates on volume-rendered images can help them detect more aneurysms.

PURPOSE: The performance of computer-aided detection (CAD) software depends on the quality and quantity of the dataset used for machine learning. If the data characteristics in development and practical use are different, the performance of CAD software degrades. In this study, we investigated changes in detection performance due to differences in training data for cerebral aneurysm detection software in head magnetic resonance angiography images. MATERIALS AND METHODS: We utilized three types of CAD software for cerebral aneurysm detection in MRA images, which were based on 3D local intensity structure analysis, graph-based features, and convolutional neural network. For each type of CAD software, we compared three types of training pattern, which were two types of training using single-site data and one type of training using multisite data. We also carried out internal and external evaluations. RESULTS: In training using single-site data, the performance of CAD software largely and unpredictably fluctuated when the training dataset was changed. Training using multisite data did not show the lowest performance among the three training patterns for any CAD software and dataset. CONCLUSION: The training of cerebral aneurysm detection software using data collected from multiple sites is desirable to ensure the stable performance of the software.

PURPOSE: To evaluate the spatial patterns of missed lung nodules in a real-life routine screening environment. MATERIALS AND METHODS: In a screening institute, 4,822 consecutive adults underwent chest CT, and each image set was independently interpreted by two radiologists in three steps: (1) independently interpreted without computer-assisted detection (CAD) software, (2) independently referred to the CAD results, (3) determined by the consensus of the two radiologists. The locations of nodules and the detection performance data were semi-automatically collected using a CAD server integrated into the reporting system. Fisher's exact test was employed for evaluating findings in different lung divisions. Probability maps were drawn to illustrate the spatial distribution of radiologists' missed nodules. RESULTS: Radiologists significantly tended to miss lung nodules in the bilateral hilar divisions (p < 0.01). Some radiologists had their own spatial pattern of missed lung nodules. CONCLUSION: Radiologists tend to miss lung nodules present in the hilar regions significantly more often than in the rest of the lung.

The purposes of this study are to propose an unsupervised anomaly detection method based on a deep neural network (DNN) model, which requires only normal images for training, and to evaluate its performance with a large chest radiograph dataset. We used the auto-encoding generative adversarial network (α-GAN) framework, which is a combination of a GAN and a variational autoencoder, as a DNN model. A total of 29,684 frontal chest radiographs from the Radiological Society of North America Pneumonia Detection Challenge dataset were used for this study (16,880 male and 12,804 female patients; average age, 47.0 years). All these images were labeled as "Normal," "No Opacity/Not Normal," or "Opacity" by board-certified radiologists. About 70% (6,853/9,790) of the Normal images were randomly sampled as the training dataset, and the rest were randomly split into the validation and test datasets in a ratio of 1:2 (7,610 and 15,221). Our anomaly detection system could correctly visualize various lesions including a lung mass, cardiomegaly, pleural effusion, bilateral hilar lymphadenopathy, and even dextrocardia. Our system detected the abnormal images with an area under the receiver operating characteristic curve (AUROC) of 0.752. The AUROCs for the abnormal labels Opacity and No Opacity/Not Normal were 0.838 and 0.704, respectively. Our DNN-based unsupervised anomaly detection method could successfully detect various diseases or anomalies in chest radiographs by training with only the normal images.

RATIONALE AND OBJECTIVES: A more accurate lung nodule detection algorithm is needed. We developed a modified three-dimensional (3D) U-net deep-learning model for the automated detection of lung nodules on chest CT images. The purpose of this study was to evaluate the accuracy of the developed modified 3D U-net deep-learning model. MATERIALS AND METHODS: In this Health Insurance Portability and Accountability Act-compliant, Institutional Review Board-approved retrospective study, the 3D U-net based deep-learning model was trained using the Lung Image Database Consortium and Image Database Resource Initiative dataset. For internal model validation, we used 89 chest CT scans that were not used for model training. For external model validation, we used 450 chest CT scans taken at an urban university hospital in Japan. Each case included at least one nodule of >5 mm identified by an experienced radiologist. We evaluated model accuracy using the competition performance metric (CPM) (average sensitivity at 1/8, 1/4, 1/2, 1, 2, 4, and 8 false-positives per scan). The 95% confidence interval (CI) was computed by bootstrapping 1000 times. RESULTS: In the internal validation, the CPM was 94.7% (95% CI: 89.1%-98.6%). In the external validation, the CPM was 83.3% (95% CI: 79.4%-86.1%). CONCLUSION: The modified 3D U-net deep-learning model showed high performance in both internal and external validation.

PURPOSE: To build a novel, open-source, purely web-based platform system to address problems in the development and clinical use of computer-assisted detection/diagnosis (CAD) software. The new platform system will replace the existing system for the development and validation of CAD software, Clinical Infrastructure for Radiologic Computation of United Solutions (CIRCUS). METHODS: In our new system, the two top-level applications visible to users are the web-based image database (CIRCUS DB; database) and the Docker plug-in-based CAD execution platform (CIRCUS CS; clinical server). These applications are built on top of a shared application programming interface server, a three-dimensional image viewer component, and an image repository. RESULTS: We successfully installed our new system into a Linux server at two clinical sites. A total of 1954 cases were registered in CIRCUS DB. We have been utilizing CIRCUS CS with four Docker-based CAD plug-ins. CONCLUSIONS: We have successfully built a new version of the CIRCUS system. Our platform was successfully implemented at two clinical sites, and we plan to publish it as an open-source software project.

Muography is a novel method of visualizing the internal structures of active volcanoes by using high-energy near-horizontally arriving cosmic muons. The purpose of this study is to show the feasibility of muography to forecast the eruption event with the aid of the convolutional neural network (CNN). In this study, seven daily consecutive muographic images were fed into the CNN to compute the probability of eruptions on the eighth day, and our CNN model was trained by hyperparameter tuning with the Bayesian optimization algorithm. By using the data acquired in Sakurajima volcano, Japan, as an example, the forecasting performance achieved a value of 0.726 for the area under the receiver operating characteristic curve, showing the reasonable correlation between the muographic images and eruption events. Our result suggests that muography has the potential for eruption forecasting of volcanoes.

In general, adversarial perturbations superimposed on inputs are realistic threats for a deep neural network (DNN). In this paper, we propose a practical generation method of such adversarial perturbation to be applied to black-box attacks that demand access to an input-output relationship only. Thus, the attackers generate such perturbation without invoking inner functions and/or accessing the inner states of a DNN. Unlike the earlier studies, the algorithm to generate the perturbation presented in this study requires much fewer query trials. Moreover, to show the effectiveness of the adversarial perturbation extracted, we experiment with a DNN for semantic segmentation. The result shows that the network is easily deceived with the perturbation generated than using uniformly distributed random noise with the same magnitude.

Preventing the oversight of anomalies in chest X-ray radiographs (CXRs) during diagnosis is a crucial issue. Deep learning (DL)-based anomaly detection methods are rapidly growing in popularity, and provide effective solutions to the problem, but the workload in labeling CXRs during the training procedure remains heavy. To reduce the workload, a novel anomaly detection method for CXRs based on weakly supervised DL is presented in this study. The DL is based on a flow-based deep neural network (DNN) framework with which two normality metrics (logarithm likelihood and logarithm likelihood ratio) can be calculated. With this method, only one set of normal CXRs requires labeling to train the DNN, then the normality of any unknown CXR can be evaluated. The area under the receiver operation characteristic curve acquired with the logarithm likelihood ratio metric ($\approx0.783$) was greater than that obtained with the logarithm likelihood metric, and was a value comparable to those in previous studies where other weakly supervised DNNs were implemented.

PURPOSE: A novel image feature set named histogram of triangular paths in graph (HoTPiG) is presented. The purpose of this study is to evaluate the feasibility of the proposed HoTPiG feature set through two clinical computer-aided detection tasks: nodule detection in lung CT images and aneurysm detection in head MR angiography images. METHODS: The HoTPiG feature set is calculated from an undirected graph structure derived from a binarized volume. The features are derived from a 3-D histogram in which each bin represents a triplet of shortest path distances between the target node and all possible node pairs near the target node. First, the vessel structure is extracted from CT/MR volumes. Then, a graph structure is extracted using an 18-neighbor rule. Using this graph, a HoTPiG feature vector is calculated at every foreground voxel. After explicit feature mapping with an exponential-χ2 kernel, each voxel is judged by a linear support vector machine classifier. The proposed method was evaluated using 300 CT and 300 MR datasets. RESULTS: The proposed method successfully detected lung nodules and cerebral aneurysms. The sensitivity was about 80% when the number of false positives was three per case for both applications. CONCLUSIONS: The HoTPiG image feature set was presented, and its high general versatility was shown through two medical lesion detection applications.

PURPOSE: The purpose of this study was to determine whether temporal subtraction (TS) computed tomography (CT) contributes to the detection of vertebral bone metastases. METHOD: The calculation of TS CT was composed of bony landmark detection, bone segmentation with a multiatlas-based method, and spatial registration. Temporal increase and decrease of the CT values were visualized in blue and red, respectively. Paired CT images of 20 patients with cancer and newly-developed vertebral metastases were analyzed. Control CT examinations of 20 different patients were also included. The presence of vertebral metastases on the TS CT was evaluated by two board-certified radiologists. Five additional board-certified radiologists and five radiology residents independently interpreted the 40 paired CT images with and without TS CT. RESULTS: In the lesion conspicuity evaluation, 96% of vertebral metastases were scored as excellent or good. In the image interpretation examination, according to free-response receiver operating characteristics analysis, the overall figure of merit (FOM) of the board-certified radiologist group was 0.892 and 0.898 with and without TS CT, respectively. The FOM of the resident group improved from 0.849 to 0.902 with viewing TS CT. In the sub-analysis focusing on the location of the lesion, the FOM of the resident group significantly improved from 0.75 to 0.92 in vertebral arch lesions (p = 0.001). CONCLUSIONS: The TS CT may be useful to detect vertebral metastases because almost all the vertebral metastases were shown to be favorable visualization. The TS CT was proven to be especially helpful for radiology residents in detecting vertebral arch metastases.

PURPOSE: Gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid (Gd-EOB-DTPA)-enhanced magnetic resonance imaging (MRI) tends to show higher diagnostic accuracy than other modalities. There is a demand for computer-assisted detection (CAD) software for Gd-EOB-DTPA-enhanced MRI. Segmentation with high accuracy is important for CAD software. We propose a liver segmentation method for Gd-EOB-DTPA-enhanced MRI that is based on a four-dimensional (4D) fully convolutional residual network (FC-ResNet). The aims of this study are to determine the best combination of an input image and output image in our proposed method and to compare our proposed method with the previous rule-based segmentation method. METHODS: We prepared a five-phase image set and a hepatobiliary phase image set as the input image sets to determine the best input image set. We also prepared a labeled liver image and labeled liver and labeled body trunk images as the output image sets to determine the best output image set. In addition, we optimized the hyperparameters of our proposed model. We used 30 cases to train our model, 10 cases to determine the hyperparameters of our model, and 20 cases to evaluate our model. RESULTS: Our network with the five-phase image set and the output image set of labeled liver and labeled body trunk images showed the highest accuracy. Our proposed method showed higher accuracy than the previous rule-based segmentation method. The Dice coefficient of the liver region was 0.944 ± 0.018. CONCLUSION: Our proposed 4D FC-ResNet showed satisfactory performance for liver segmentation as preprocessing in CAD software.

BACKGROUND: This retrospective study examined the performance of computer-assisted detection in the identification of pulmonary metastases. METHODS: Fifty-five patients (41.8% male) who underwent surgery for metastatic lung tumors in our hospital from 2005 to 2012 were included. Computer-assisted detection software configured to display the top five nodule candidates according to likelihood was applied as the first reader for the preoperative computed tomography images. Results from the software were classified as "metastatic nodule", "benign nodule", or "false-positive finding" by two observers. RESULTS: Computer-assisted detection identified 85.3% (64/75) of pulmonary metastases that radiologists had detected, and 3 more (4%, 3/75) that radiologists had overlooked. Nodule candidates identified by computer-assisted detection included 86 benign nodules (median size 3.1 mm, range 1.2-18.7 mm) and 121 false-positive findings. CONCLUSIONS: Computer-assisted detection identified pulmonary metastases overlooked by radiologists. However, this was at the cost of identifying a substantial number of benign nodules and false-positive findings.

PURPOSE: For the development of computer-assisted detection (CAD) software using voxel-based classification, gold standards defined by pixel-by-pixel painting, called painted gold standards, are desirable. However, for radiologists who define gold standards, a simplified method of definition is desirable. One of the simplest methods of defining gold standards is a spherical region, called a spherical gold standard. In this study, we investigated whether spherical gold standards can be used as an alternative to painted gold standards for computerized detection using voxel-based classification. MATERIALS AND METHODS: The spherical gold standards were determined by the center of gravity and the maximum diameter. We compared two types of gold standard, painted gold standards and spherical gold standards, by two types of CAD software using voxel-based classification. RESULTS: The time required to paint the area of one lesion was 4.7-6.5 times longer than the time required to define a spherical gold standard. For the same performance of the CAD software, the number of training cases required for the spherical gold standard was 1.6-7.6 times that for the painted gold standards. CONCLUSION: Spherical gold standards can be used as an alternative to painted gold standards for the computerized detection of lesions with simple shapes.

© 2018 Association for Computing Machinery. The reading workload for radiologists is increasing because the numbers of examinations and images per examination are increasing due to the technical progress on imaging modalities such as computed tomography and magnetic resonance imaging. A computer-assisted detection (CAD) system based on machine learning is expected to assist radiologists. The preliminary results of a multi-institutional study indicate that the performance of the CAD system for each institution improved using training data of other institutions. This indicates that transfer learning may be useful for developing the CAD systems among multiple institutions. In this paper, we focus on transfer learning without sharing training data due to the need to protect personal information in each institution. Moreover, we raise a problem of negative transfer in CAD system and propose an algorithm for inhibiting negative transfer. Our algorithm provides a theoretical guarantee for managing CAD software in terms of transfer learning and exhibits experimentally better performance compared to that of the current algorithm in cerebral aneurysm detection.