Internal Test Dataset Research Articles

Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos

Surgical video analysis facilitates education and research. However, video recordings of endoscopic surgeries can contain privacy-sensitive information, especially if the endoscopic camera is moved out of the body of patients and out-of-body scenes are recorded. Therefore, identification of out-of-body scenes in endoscopic videos is of major importance to preserve the privacy of patients and operating room staff. This study developed and validated a deep learning model for the identification of out-of-body images in endoscopic videos. The model was trained and evaluated on an internal dataset of 12 different types of laparoscopic and robotic surgeries and was externally validated on two independent multicentric test datasets of laparoscopic gastric bypass and cholecystectomy surgeries. Model performance was evaluated compared to human ground truth annotations measuring the receiver operating characteristic area under the curve (ROC AUC). The internal dataset consisting of 356,267 images from 48 videos and the two multicentric test datasets consisting of 54,385 and 58,349 images from 10 and 20 videos, respectively, were annotated. The model identified out-of-body images with 99.97% ROC AUC on the internal test dataset. Mean ± standard deviation ROC AUC on the multicentric gastric bypass dataset was 99.94 ± 0.07% and 99.71 ± 0.40% on the multicentric cholecystectomy dataset, respectively. The model can reliably identify out-of-body images in endoscopic videos and is publicly shared. This facilitates privacy preservation in surgical video analysis.

The Zhu-Lu formula: a machine learning-based intraocular lens power calculation formula for highly myopic eyes

BackgroundTo develop a novel machine learning-based intraocular lens (IOL) power calculation formula for highly myopic eyes.MethodsA total of 1828 eyes (from 1828 highly myopic patients) undergoing cataract surgery in our hospital were used as the internal dataset, and 151 eyes from 151 highly myopic patients from two other hospitals were used as external test dataset. The Zhu-Lu formula was developed based on the eXtreme Gradient Boosting and the support vector regression algorithms. Its accuracy was compared in the internal and external test datasets with the Barrett Universal II (BUII), Emmetropia Verifying Optical (EVO) 2.0, Kane, Pearl-DGS and Radial Basis Function (RBF) 3.0 formulas.ResultsIn the internal test dataset, the Zhu-Lu, RBF 3.0 and BUII ranked top three from low to high taking into account standard deviations (SDs) of prediction errors (PEs). The Zhu-Lu and RBF 3.0 showed significantly lower median absolute errors (MedAEs) than the other formulas (all P < 0.05). In the external test dataset, the Zhu-Lu, Kane and EVO 2.0 ranked top three from low to high considering SDs of PEs. The Zhu-Lu formula showed a comparable MedAE with BUII and EVO 2.0 but significantly lower than Kane, Pearl-DGS and RBF 3.0 (all P < 0.05). The Zhu-Lu formula ranked first regarding the percentages of eyes within ± 0.50 D of the PE in both test datasets (internal: 80.61%; external: 72.85%). In the axial length subgroup analysis, the PE of the Zhu-Lu stayed stably close to zero in all subgroups.ConclusionsThe novel IOL power calculation formula for highly myopic eyes demonstrated improved and stable predictive accuracy compared with other artificial intelligence-based formulas.

Densely Convolutional Spatial Attention Network for nuclei segmentation of histological images for computational pathology.

Automatic nuclear segmentation in digital microscopic tissue images can aid pathologists to extract high-quality features for nuclear morphometrics and other analyses. However, image segmentation is a challenging task in medical image processing and analysis. This study aimed to develop a deep learning-based method for nuclei segmentation of histological images for computational pathology. The original U-Net model sometime has a caveat in exploring significant features. Herein, we present the Densely Convolutional Spatial Attention Network (DCSA-Net) model based on U-Net to perform the segmentation task. Furthermore, the developed model was tested on external multi-tissue dataset - MoNuSeg. To develop deep learning algorithms for well-segmenting nuclei, a large quantity of data are mandatory, which is expensive and less feasible. We collected hematoxylin and eosin-stained image data sets from two hospitals to train the model with a variety of nuclear appearances. Because of the limited number of annotated pathology images, we introduced a small publicly accessible data set of prostate cancer (PCa) with more than 16,000 labeled nuclei. Nevertheless, to construct our proposed model, we developed the DCSA module, an attention mechanism for capturing useful information from raw images. We also used several other artificial intelligence-based segmentation methods and tools to compare their results to our proposed technique. To prioritize the performance of nuclei segmentation, we evaluated the model's outputs based on the Accuracy, Dice coefficient (DC), and Jaccard coefficient (JC) scores. The proposed technique outperformed the other methods and achieved superior nuclei segmentation with accuracy, DC, and JC of 96.4% (95% confidence interval [CI]: 96.2 - 96.6), 81.8 (95% CI: 80.8 - 83.0), and 69.3 (95% CI: 68.2 - 70.0), respectively, on the internal test data set. Our proposed method demonstrates superior performance in segmenting cell nuclei of histological images from internal and external datasets, and outperforms many standard segmentation algorithms used for comparative analysis.

Characterization of high-grade prostate cancer at multiparametric MRI using a radiomic-based computer-aided diagnosis system as standalone and second reader

PurposeThe purpose of this study was to develop and test across various scanners a zone-specific region-of-interest (ROI)-based computer-aided diagnosis system (CAD) aimed at characterizing, on MRI, International Society of Urological Pathology (ISUP) grade≥2 prostate cancers. Materials and methodsROI-based quantitative models were selected in multi-vendor training (265 pre-prostatectomy MRIs) and pre-test (112 pre-biopsy MRIs) datasets. The best peripheral and transition zone models were combined and retrospectively assessed in internal (158 pre-biopsy MRIs) and external (104 pre-biopsy MRIs) test datasets. Two radiologists (R1/R2) retrospectively delineated the lesions targeted at biopsy in test datasets. The CAD area under the receiver operating characteristic curve (AUC) for characterizing ISUP≥2 cancers was compared to that of the Prostate Imaging-Reporting and Data System version2 (PI-RADSv2) score prospectively assigned to targeted lesions. ResultsThe best models used the 25th apparent diffusion coefficient (ADC) percentile in transition zone and the 2nd ADC percentile and normalized wash-in rate in peripheral zone. The PI-RADSv2 AUCs were 82% (95% confidence interval [CI]: 74–87) and 86% (95% CI: 81–91) in the internal and external test datasets respectively. They were not different from the CAD AUCs obtained with R1 and R2 delineations, in the internal (82% [95% CI: 76–89], P = 0.95 and 85% [95% CI: 78–91], P = 0.55) and external (82% [95% CI: 74–91], P = 0.41 and 86% [95% CI:78–95], P = 0.98) test datasets. The CAD yielded sensitivities of 86–89% and 90–91%, and specificities of 64–65% and 69–75% in the internal and external test datasets respectively. ConclusionThe CAD performance for characterizing ISUP grade≥2 prostate cancers on MRI is not different from that of PI-RADSv2 score across two test datasets.

A Deep Learning Pipeline Using Prior Knowledge for Automatic Evaluation of Placenta Accreta Spectrum Disorders With MRI.

The diagnosis of prenatal placenta accreta spectrum (PAS) with magnetic resonance imaging (MRI) is highly dependent on radiologists' experience. A deep learning (DL) method using the prior knowledge that PAS-related signs are generally found along the utero-placental borderline (UPB) may help radiologists, especially those with less experience, to mitigate this issue. To develop a DL tool for antenatal diagnosis of PAS using T2-weighted MR images. Retrospective. Five hundred and forty pregnant women with clinically suspected PAS disorders from two institutions, divided into training (409), internal test (103), and external test (28) datasets. Sagittal T2-weighted fast spin echo sequence at 1.5 T and 3 T. An nnU-Net was trained for placenta segmentation. The UPB straightening approach was used to extract the utero-placental boundary region. The UPB image was then fed into DenseNet-PAS for PAS diagnosis. DenseNet-PP learnt placental position information to improve the PAS diagnosis performance. Three radiologists with 8, 10, and 12 years of experience independently evaluated the images. Two radiologists marked the placenta tissue. Histopathological findings were the reference standard. Area under the curve (AUC) was used to evaluate the classification. Dice coefficient evaluated the segmentation between radiologists and the model performance. The Mann-Whitney U-test or the chi-squared test assessed the significance of differences. Decision curve analysis was used to determine clinical effectiveness. DeLong's test was used to compare AUCs. Of the 540 patients, 170 had PAS disorders confirmed by histopathology. The DL model using UPB images and placental position yielded the highest AUC of 0.860 and 0.897 in internal test and external test cohorts, respectively, significantly exceeding the performance of three radiologists (internal test AUC, 0.737-0.770). By extracting the UPB image, this fully automatic DL pipeline achieved high accuracy and may assist radiologists in PAS diagnosis using MRI. 3 TECHNICAL EFFICACY: Stage 2.

DH-452784-4 AN ELECTROCARDIOGRAPHY-BASED DEEP LEARNING ALGORITHM FOR IDENTIFICATION OF INDIVIDUALS AT INCREASED RISK OF SUDDEN CARDIAC ARREST

Sudden cardiac arrest (SCA) is a major public health problem with significant need to improve screening methods for detection of high risk. Recently developed ECG-based deep learning (DL) models have shown high accuracy for detection of selected heart disease conditions. To train and validate a 12-lead ECG-based DL algorithm for detection of increased SCA risk. Individuals who suffered out-of-hospital SCA were prospectively ascertained in the Portland, Oregon, metro area (The Oregon Sudden Unexpected Death Study, catchment population ∼1 million, 2002-onwards). A total of 1,838 12-lead ECGs from 1,796 cases obtained by health care providers prior and unrelated to the individual SCA events were digitized and employed to develop an ECG-based DL model for prediction of SCA risk (training, validation, and testing). External validation was performed in 717 ECGs from 714 SCA cases (The Ventura Prediction of Sudden Death in Multi-ethnic Communities study, catchment population ≈850,000, 2015-onwards). Two separate control group samples were obtained from 1,350 ECGs taken from 1,325 clinically matched individuals. Compared to SCA cases in the internal dataset, individuals in the external dataset were older, and more often female or Hispanic. The prevalence of prior myocardial infarction and heart failure was lower among SCA cases in the external cohort. The DL model achieved an AUC of 0.887 (95% CI 0.857-0.915) for detection of SCA cases in the internal test dataset compared to control patients in the held-out test cohort. Despite the demographic and clinical differences, the model was successfully validated among external SCA cases with an AUC of 0.821 (95% CI 0.791-0.850). The DL model performed significantly better than a previously developed and validated conventional ECG electrical risk score, which achieved an AUC of 0.705 (0.661-0.749) in the internal cohort and 0.742 (0.710-0.774) in the external cohort. An ECG-based DL model distinguished SCA cases from clinically matched control subjects with high accuracy, performing better than a conventional ECG risk model. These findings suggest that this DL model has potential for further evaluation as a screening tool for SCA risk.

Multi-channel deep learning model-based myocardial spatial–temporal morphology feature on cardiac MRI cine images diagnoses the cause of LVH

BackgroundTo develop a fully automatic framework for the diagnosis of cause for left ventricular hypertrophy (LVH) via cardiac cine images.MethodsA total of 302 LVH patients with cine MRI images were recruited as the primary cohort. Another 53 LVH patients prospectively collected or from multi-centers were used as the external test dataset. Different models based on the cardiac regions (Model 1), segmented ventricle (Model 2) and ventricle mask (Model 3) were constructed. The diagnostic performance was accessed by the confusion matrix with respect to overall accuracy. The capability of the predictive models for binary classification of cardiac amyloidosis (CA), hypertrophic cardiomyopathy (HCM) or hypertensive heart disease (HHD) were also evaluated. Additionally, the diagnostic performance of best Model was compared with that of 7 radiologists/cardiologists.ResultsModel 3 showed the best performance with an overall classification accuracy up to 77.4% in the external test datasets. On the subtasks for identifying CA, HCM or HHD only, Model 3 also achieved the best performance with AUCs yielding 0.895–0.980, 0.879–0.984 and 0.848–0.983 in the validation, internal test and external test datasets, respectively. The deep learning model showed non-inferior diagnostic capability to the cardiovascular imaging expert and outperformed other radiologists/cardiologists.ConclusionThe combined model based on the mask of left ventricular segmented from multi-sequences cine MR images shows favorable and robust performance in diagnosing the cause of left ventricular hypertrophy, which could be served as a noninvasive tool and help clinical decision.

Evaluation of a computer-aided diagnostic model for corneal diseases by analyzing in vivo confocal microscopy images.

In order to automatically and rapidly recognize the layers of corneal images using in vivo confocal microscopy (IVCM) and classify them into normal and abnormal images, a computer-aided diagnostic model was developed and tested based on deep learning to reduce physicians' workload. A total of 19,612 corneal images were retrospectively collected from 423 patients who underwent IVCM between January 2021 and August 2022 from Renmin Hospital of Wuhan University (Wuhan, China) and Zhongnan Hospital of Wuhan University (Wuhan, China). Images were then reviewed and categorized by three corneal specialists before training and testing the models, including the layer recognition model (epithelium, bowman's membrane, stroma, and endothelium) and diagnostic model, to identify the layers of corneal images and distinguish normal images from abnormal images. Totally, 580 database-independent IVCM images were used in a human-machine competition to assess the speed and accuracy of image recognition by 4 ophthalmologists and artificial intelligence (AI). To evaluate the efficacy of the model, 8 trainees were employed to recognize these 580 images both with and without model assistance, and the results of the two evaluations were analyzed to explore the effects of model assistance. The accuracy of the model reached 0.914, 0.957, 0.967, and 0.950 for the recognition of 4 layers of epithelium, bowman's membrane, stroma, and endothelium in the internal test dataset, respectively, and it was 0.961, 0.932, 0.945, and 0.959 for the recognition of normal/abnormal images at each layer, respectively. In the external test dataset, the accuracy of the recognition of corneal layers was 0.960, 0.965, 0.966, and 0.964, respectively, and the accuracy of normal/abnormal image recognition was 0.983, 0.972, 0.940, and 0.982, respectively. In the human-machine competition, the model achieved an accuracy of 0.929, which was similar to that of specialists and higher than that of senior physicians, and the recognition speed was 237 times faster than that of specialists. With model assistance, the accuracy of trainees increased from 0.712 to 0.886. A computer-aided diagnostic model was developed for IVCM images based on deep learning, which rapidly recognized the layers of corneal images and classified them as normal and abnormal. This model can increase the efficacy of clinical diagnosis and assist physicians in training and learning for clinical purposes.

Multi-arm U-Net with dense input and skip connectivity for T2 lesion segmentation in clinical trials of multiple sclerosis

T2 lesion quantification plays a crucial role in monitoring disease progression and evaluating treatment response in multiple sclerosis (MS). We developed a 3D, multi-arm U-Net for T2 lesion segmentation, which was trained on a large, multicenter clinical trial dataset of relapsing MS. We investigated its generalization to other relapsing and primary progressive MS clinical trial datasets, and to an external dataset from the MICCAI 2016 MS lesion segmentation challenge. Additionally, we assessed the model’s ability to reproduce the separation of T2 lesion volumes between treatment and control arms; and the association of baseline T2 lesion volumes with clinical disability scores compared with manual lesion annotations. The trained model achieved a mean dice coefficient of ≥ 0.66 and a lesion detection sensitivity of ≥ 0.72 across the internal test datasets. On the external test dataset, the model achieved a mean dice coefficient of 0.62, which is comparable to 0.59 from the best model in the challenge, and a lesion detection sensitivity of 0.68. Lesion detection performance was reduced for smaller lesions (≤ 30 μL, 3–10 voxels). The model successfully maintained the separation of the longitudinal changes in T2 lesion volumes between the treatment and control arms. Such tools could facilitate semi-automated MS lesion quantification; and reduce rater burden in clinical trials.

Development and validation of machine learning models to predict survival of patients with resected stage-III NSCLC.

To compare the performance of three machine learning algorithms with the tumor, node, and metastasis (TNM) staging system in survival prediction and validate the individual adjuvant treatment recommendations plan based on the optimal model. In this study, we trained three machine learning madel and validated 3 machine learning survival models-deep learning neural network, random forest and cox proportional hazard model- using the data of patients with stage-al3 NSCLC patients who received resection surgery from the National Cancer Institute Surveillance, Epidemiology, and End Results (SEER) database from 2012 to 2017,the performance of survival predication from all machine learning models were assessed using a concordance index (c-index) and the averaged c-index is utilized for cross-validation. The optimal model was externally validated in an independent cohort from Shaanxi Provincial People's Hospital. Then we compare the performance of the optimal model and TNM staging system. Finally, we developed a Cloud-based recommendation system for adjuvant therapy to visualize survival curve of each treatment plan and deployed on the internet. A total of 4617 patients were included in this study. The deep learning network performed more stably and accurately in predicting stage-iii NSCLC resected patients survival than the random survival forest and Cox proportional hazard model on the internal test dataset (C-index=0.834 vs. 0.678 vs. 0.640) and better than TNM staging system (C-index=0.820 vs. 0.650) in the external validation. The individual patient who follow the reference from recommendation system had superior survival compared to those who did not. The predicted 5-year-survival curve for each adjuvant treatment plan could be accessed in the recommender system via the browser. Deep learning model has several advantages over linear model and random forest model in prognostic predication and treatment recommendations. This novel analytical approach may provide accurate predication on individual survival and treatment recommendations for resected Stage-iii NSCLC patients.

Negative Samples for Improving Object Detection-A Case Study in AI-Assisted Colonoscopy for Polyp Detection.

Deep learning object-detection models are being successfully applied to develop computer-aided diagnosis systems for aiding polyp detection during colonoscopies. Here, we evidence the need to include negative samples for both (i) reducing false positives during the polyp-finding phase, by including images with artifacts that may confuse the detection models (e.g., medical instruments, water jets, feces, blood, excessive proximity of the camera to the colon wall, blurred images, etc.) that are usually not included in model development datasets, and (ii) correctly estimating a more realistic performance of the models. By retraining our previously developed YOLOv3-based detection model with a dataset that includes 15% of additional not-polyp images with a variety of artifacts, we were able to generally improve its F1 performance in our internal test datasets (from an average F1 of 0.869 to 0.893), which now include such type of images, as well as in four public datasets that include not-polyp images (from an average F1 of 0.695 to 0.722).

Machine learning-based prediction of intraoperative hypoxemia for pediatric patients.

Reducing the duration of intraoperative hypoxemia in pediatric patients by means of rapid detection and early intervention is considered crucial by clinicians. We aimed to develop and validate a machine learning model that can predict intraoperative hypoxemia events 1 min ahead in children undergoing general anesthesia. This retrospective study used prospectively collected intraoperative vital signs and parameters from the anesthesia ventilator machine extracted every 2 s in pediatric patients undergoing surgery under general anesthesia between January 2019 and October 2020 in a tertiary academic hospital. Intraoperative hypoxemia was defined as oxygen saturation <95% at any point during surgery. Three common machine learning techniques were employed to develop models using the training dataset: gradient-boosting machine (GBM), long short-term memory (LSTM), and transformer. The performances of the models were compared using the area under the receiver operating characteristics curve using randomly assigned internal testing dataset. We also validated the developed models using temporal holdout dataset. Pediatric patient surgery cases between November 2020 and January 2021 were used. The performances of the models were compared using the area under the receiver operating characteristic curve (AUROC). In total, 1,540 (11.73%) patients with intraoperative hypoxemia out of 13,130 patients' records with 2,367 episodes were included for developing the model dataset. After model development, 200 (13.25%) of the 1,510 patients' records with 289 episodes were used for holdout validation. Among the models developed, the GBM had the highest AUROC of 0.904 (95% confidence interval [CI] 0.902 to 0.906), which was significantly higher than that of the LSTM (0.843, 95% CI 0.840 to 0.846 P < .001) and the transformer model (0.885, 95% CI, 0.882-0.887, P < .001). In holdout validation, GBM also demonstrated best performance with an AUROC of 0.939 (95% CI 0.936 to 0.941) which was better than LSTM (0.904, 95% CI 0.900 to 0.907, P < .001) and the transformer model (0.929, 95% CI 0.926 to 0.932, P < .001). Machine learning models can be used to predict upcoming intraoperative hypoxemia in real-time based on the biosignals acquired by patient monitors, which can be useful for clinicians for prediction and proactive treatment of hypoxemia in an intraoperative setting.

A Deep Learning Approach for Automated Bone Removal from Computed Tomography Angiography of the Brain.

Advanced visualization techniques such as maximum intensity projection (MIP) and volume rendering (VR) are useful for evaluating neurovascular anatomy on CT angiography (CTA) of the brain; however, interference from surrounding osseous anatomy is common. Existing methods for removing bone from CTA images are limited in scope and/or accuracy, particularly at the skull base. We present a new brain CTA bone removal tool, which addresses many of these limitations.A deep convolutional neural network was designed and trained for bone removal using 72 brain CTAs. The model was tested on 15 CTAs from the same data source and 17 CTAs from an independent external dataset. Bone removal accuracy was assessed quantitatively, by comparing automated segmentation results to manual segmentations, and qualitatively by evaluating VR visualization of the carotid siphons compared to an existing method for automated bone removal.Average Dice overlap between automated and manual segmentations from the internal and external test datasets were 0.986 and 0.979 respectively. This was superior compared to a publicly available noncontrast head CT bone removal algorithm which had a Dice overlap of 0.947 (internal dataset) and 0.938 (external dataset). Our algorithm yielded better VR visualization of the carotid siphons than the publicly available bone removal tool in 14 out of 15 CTAs (93%, chi-square statistic of 22.5, p-value of < 0.00001) from the internal test dataset and 15 out of 17 CTAs (88%, chi-square statistic of 23.1, p-value of < 0.00001) from the external test dataset. Bone removal allowed subjectively superior MIP and VR visualization of vascular anatomy/pathology.The proposed brain CTA bone removal algorithm is rapid, accurate, and allows superior visualization of vascular anatomy and pathology compared to other available techniques and was validated on an independent external dataset.

P254 Deep Learning Model for Distinguishing MES 0 and MES 1 in Patients with Ulcerative Colitis

Abstract Background Endoscopic remission has recently been defined as a Mayo endoscopic sub-score (MES) of 0. Therefore, patients with MES 1 need to step up to achieve an MES of 0 on the endoscopic score. In discriminating between MES 0 and 1, the inter-observer variation was very severe among the endoscopists. This study aimed to narrow the gap in distinguishing between MES 0 and MES1 using a deep learning model. Methods From the endoscopic images of 492 ulcerative colitis (UC) patients with MES improvement (MES 0 or MES 1) from January 2018 to December 2019 at this center, two representative images of the colon and rectum were selected, and a total of 984 images were analyzed. Our model was composed of a convolutional neural network (CNN)-based encoder, two auxiliary classifiers for the colon and rectum, and a final MES classifier for the combined image features of the two inputs. The following three experiments were conducted to validate and test the proposed model: 12-fold cross-validation for representative model selection, performance comparison with a novice group through an internal test dataset, and an external test with Hyperkvasir, a public gastrointestinal endoscopic image dataset. Results In the internal test, our model showed an F1-score of 0.92. This was, on average, 0.11 higher than each of the seven novices and 0.02 higher than their consensus. Considering MES 1 as positive, the area under the receiver operating characteristic curve (AUROC) was 0.97, and the area under the precision-recall curve (AUPRC) was 0.98. In the external test with Hyperkvasir, our model showed an F1-score of 0.89, AUROC of 0.86, and AUPRC of 0.97. Conclusion We found that the proposed CNN-based model that integrates the image features of the colon and rectum has superior performance in discriminating between MES 0 and MES 1 in patients with UC. Further prospective studies are required to prove the clinical utility of our model in clinical practice.

Deep Learning for Noninvasive Assessment of H3 K27M Mutation Status in Diffuse Midline Gliomas Using MR Imaging.

Determination of H3 K27M mutation in diffuse midline glioma (DMG) is key for prognostic assessment and stratifying patient subgroups for clinical trials. MRI can noninvasively depict morphological and metabolic characteristics of H3 K27M mutant DMG. This study aimed to develop a deep learning (DL) approach to noninvasively predict H3 K27M mutation in DMG using T2-weighted images. Retrospective and prospective. For diffuse midline brain gliomas, 341 patients from Center-1 (27 ± 19 years, 184 males), 42 patients from Center-2 (33 ± 19 years, 27 males) and 35 patients (37 ± 18 years, 24 males). For diffuse spinal cord gliomas, 133 patients from Center-1 (30 ± 15 years, 80 males). 5T and 3T, T2-weighted turbo spin echo imaging. Conventional radiological features were independently reviewed by two neuroradiologists. H3 K27M status was determined by histopathological examination. The Dice coefficient was used to evaluate segmentation performance. Classification performance was evaluated using accuracy, sensitivity, specificity, and area under the curve. Pearson's Chi-squared test, Fisher's exact test, two-sample Student's t-test and Mann-Whitney U test. A two-sided P value <0.05 was considered statistically significant. In the testing cohort, Dice coefficients of tumor segmentation using DL were 0.87 for diffuse midline brain and 0.81 for spinal cord gliomas. In the internal prospective testing dataset, the predictive accuracies, sensitivities, and specificities of H3 K27M mutation status were 92.1%, 98.2%, 82.9% in diffuse midline brain gliomas and 85.4%, 88.9%, 82.6% in spinal cord gliomas. Furthermore, this study showed that the performance generalizes to external institutions, with predictive accuracies of 85.7%-90.5%, sensitivities of 90.9%-96.0%, and specificities of 82.4%-83.3%. In this study, an automatic DL framework was developed and validated for accurately predicting H3 K27M mutation using T2-weighted images, which could contribute to the noninvasive determination of H3 K27M status for clinical decision-making. 2 Technical Efficacy: Stage 2.

DeepFundus: A flow-cytometry-like image quality classifier for boosting the whole life cycle of medical artificial intelligence

Medical artificial intelligence (AI) has been moving from the research phase to clinical implementation. However, most AI-based models are mainly built using high-quality images preprocessed in the laboratory, which is not representative of real-world settings. This dataset bias proves a major driver of AI system dysfunction. Inspired by the design of flow cytometry, DeepFundus, a deep-learning-based fundus image classifier, is developed to provide automated and multidimensional image sorting to address this data quality gap. DeepFundus achieves areas under the receiver operating characteristic curves (AUCs) over 0.9 in image classification concerning overall quality, clinical quality factors, and structural quality analysis on both the internal test and national validation datasets. Additionally, DeepFundus can be integrated into both model development and clinical application of AI diagnostics to significantly enhance model performance for detecting multiple retinopathies. DeepFundus can be used to construct a data-driven paradigm for improving the entire life cycle of medical AI practice.

Automated lung cancer assessment on 18F-PET/CT using Retina U-Net and anatomical region segmentation

ObjectivesTo develop and test a Retina U-Net algorithm for the detection of primary lung tumors and associated metastases of all stages on FDG-PET/CT.MethodsA data set consisting of 364 FDG-PET/CTs of patients with histologically confirmed lung cancer was used for algorithm development and internal testing. The data set comprised tumors of all stages. All lung tumors (T), lymphatic metastases (N), and distant metastases (M) were manually segmented as 3D volumes using whole-body PET/CT series. The data set was split into a training (n = 216), validation (n = 74), and internal test data set (n = 74). Detection performance for all lesion types at multiple classifier thresholds was evaluated and false-positive-findings-per-case (FP/c) calculated. Next, detected lesions were assigned to categories T, N, or M using an automated anatomical region segmentation. Furthermore, reasons for FPs were visually assessed and analyzed. Finally, performance was tested on 20 PET/CTs from another institution.ResultsSensitivity for T lesions was 86.2% (95% CI: 77.2–92.7) at a FP/c of 2.0 on the internal test set. The anatomical correlate to most FPs was the physiological activity of bone marrow (16.8%). TNM categorization based on the anatomical region approach was correct in 94.3% of lesions. Performance on the external test set confirmed the good performance of the algorithm (overall detection rate = 88.8% (95% CI: 82.5–93.5%) and FP/c = 2.7).ConclusionsRetina U-Nets are a valuable tool for tumor detection tasks on PET/CT and can form the backbone of reading assistance tools in this field. FPs have anatomical correlates that can lead the way to further algorithm improvements. The code is publicly available.Key Points• Detection of malignant lesions in PET/CT with Retina U-Net is feasible.• All false-positive findings had anatomical correlates, physiological bone marrow activity being the most prevalent.• Retina U-Nets can build the backbone for tools assisting imaging professionals in lung tumor staging.

IERM: An Interpretable Deep Learning System to Classify Epiretinal Membrane for Different Optical Coherence Tomography Devices: A Multi-Center Analysis.

Background: Epiretinal membranes (ERM) have been found to be common among individuals >50 years old. However, the severity grading assessment for ERM based on optical coherence tomography (OCT) images has remained a challenge due to lacking reliable and interpretable analysis methods. Thus, this study aimed to develop a two-stage deep learning (DL) system named iERM to provide accurate automatic grading of ERM for clinical practice. Methods: The iERM was trained based on human segmentation of key features to improve classification performance and simultaneously provide interpretability to the classification results. We developed and tested iERM using a total of 4547 OCT B-Scans of four different commercial OCT devices that were collected from nine international medical centers. Results: As per the results, the integrated network effectively improved the grading performance by 1−5.9% compared with the traditional classification DL model and achieved high accuracy scores of 82.9%, 87.0%, and 79.4% in the internal test dataset and two external test datasets, respectively. This is comparable to retinal specialists whose average accuracy scores are 87.8% and 79.4% in two external test datasets. Conclusion: This study proved to be a benchmark method to improve the performance and enhance the interpretability of the traditional DL model with the implementation of segmentation based on prior human knowledge. It may have the potential to provide precise guidance for ERM diagnosis and treatment.

MiDruglikeness: Subdivisional Drug-Likeness Prediction Models Using Active Ensemble Learning Strategies.

The drug development pipeline involves several stages including in vitro assays, in vivo assays, and clinical trials. For candidate selection, it is important to consider that a compound will successfully pass through these stages. Using graph neural networks, we developed three subdivisional models to individually predict the capacity of a compound to enter in vivo testing, clinical trials, and market approval stages. Furthermore, we proposed a strategy combing both active learning and ensemble learning to improve the quality of the models. The models achieved satisfactory performance in the internal test datasets and four self-collected external test datasets. We also employed the models as a general index to make an evaluation on a widely known benchmark dataset DEKOIS 2.0, and surprisingly found a powerful ability on virtual screening tasks. Our model system (termed as miDruglikeness) provides a comprehensive drug-likeness prediction tool for drug discovery and development.

Artificial intelligence for detecting and delineating the extent of superficial esophageal squamous cell carcinoma and precancerous lesions under narrow-band imaging (with video)

Although narrow-band imaging (NBI) is a useful modality for detecting and delineating esophageal squamous cell carcinoma (ESCC), there is a risk of incorrectly determining the margins of some lesions even with NBI. This study aimed to develop an artificial intelligence (AI) system for detecting superficial ESCC and precancerous lesions and delineating the extent of lesions under NBI. Nonmagnified NBI images from 4 hospitals were collected and annotated. Internal and external image test datasets were used to evaluate the detection and delineation performance of the system. The delineation performance of the system was compared with that of endoscopists. Furthermore, the system was directly integrated into the endoscopy equipment, and its real-time diagnostic capability was prospectively estimated. The system was trained and tested using 10,047 still images and 140 videos from 1112 patients and 1183 lesions. In the image testing, the accuracy of the system in detecting lesions in internal and external tests was 92.4% and 89.9%, respectively. The accuracy of the system in delineating extents in internal and external tests was 88.9% and 87.0%, respectively. The delineation performance of the system was superior to that of junior endoscopists and similar to that of senior endoscopists. In the prospective clinical evaluation, the system exhibited satisfactory performance, with an accuracy of 91.4% in detecting lesions and an accuracy of 85.9% in delineating extents. The proposed AI system could accurately detect superficial ESCC and precancerous lesions and delineate the extent of lesions under NBI.