Deep Learning System Research Articles

Cross-Modal Retrieval Between 13C NMR Spectra and Structures Based on Focused Libraries.

Library matching by comparing carbon-13 nuclear magnetic resonance (13C NMR) spectra with spectral data in the library is a crucial method for compound identification. In our previous paper, we introduced a deep contrastive learning system called CReSS, which used a library that contained more structures. However, CReSS has two limitations: there were no unknown structures in the library, and a redundant library reduces the structure-elucidation accuracy. Herein, we replaced the oversize traditional libraries with focused libraries containing a small number of molecules. A previously generative model, CMGNet, was used to generate focused libraries for CReSS. The combined model achieved a Top-10 accuracy of 54.03% when tested on 6,471 13C NMR spectra. In comparison, CReSS with a random reference structure library achieved an accuracy of only 9.17%. Furthermore, to expand the advantages of the focused libraries, we proposed SAmpRNN, which is a recurrent neural network (RNN). With the large focused library amplified by SAmpRNN, the structure-identification accuracy of the model increased in 70.0% of the 30 random example cases. In general, cross-modal retrieval between 13C NMR spectra and structures based on focused libraries (CFLS) achieved high accuracy and provided more accurate candidate structures than traditional libraries for compound identification.

Deep Learning-Based Automated Detection of Retinal Breaks and Detachments on Fundus Photography.

The purpose of this study was to develop a deep learning algorithm, to detect retinal breaks and retinal detachments on ultra-widefield fundus (UWF) optos images using artificial intelligence (AI). Optomap UWF images of the database were annotated to four groups by two retina specialists: (1) retinal breaks without detachment, (2) retinal breaks with retinal detachment, (3) retinal detachment without visible retinal breaks, and (4) a combination of groups 1 to 3. The fundus image data set was split into a training set and an independent test set following an 80% to 20% ratio. Image preprocessing methods were applied. An EfficientNet classification model was trained with the training set and evaluated with the test set. A total of 2489 UWF images were included into the dataset, resulting in a training set size of 2008 UWF images and a test set size of 481 images. The classification models achieved an area under the receiver operating characteristic curve (AUC) on the testing set of 0.975 regarding lesion detection, an AUC of 0.972 for retinal detachment and an AUC of 0.913 for retinal breaks. A deep learning system to detect retinal breaks and retinal detachment using UWF images is feasible and has a good specificity. This is relevant for clinical routine as there can be a high rate of missed breaks in clinics. Future clinical studies will be necessary to evaluate the cost-effectiveness of applying such an algorithm as an automated auxiliary tool in a large practices or tertiary referral centers. This study demonstrates the relevance of applying AI in diagnosing peripheral retinal breaks in clinical routine in UWF fundus images.

GAN-Based Motion Artifact Correction of 3D MR Volumes Using an Image-to-Image Translation Algorithm.

The quality of brain MRI volumes is often compromised by motion artifacts arising from intricate respiratory patterns and involuntary head movements, manifesting as blurring and ghosting that markedly degrade imaging quality. In this study, we introduce an innovative approach employing a 3D deep learning framework to restore brain MR volumes afflicted by motion artifacts. The framework integrates a densely connected 3D U-net architecture augmented by generative adversarial network (GAN)-informed training with a novel volumetric reconstruction loss function tailored to 3D GAN to enhance the quality of the volumes. Our methodology is substantiated through comprehensive experimentation involving a diverse set of motion artifact-affected MR volumes. The generated high-quality MR volumes have similar volumetric signatures comparable to motion-free MR volumes after motion correction. This underscores the significant potential of harnessing this 3D deep learning system to aid in the rectification of motion artifacts in brain MR volumes, highlighting a promising avenue for advanced clinical applications.

Modular neural network for Edge-based Detection of early-stage IoT Botnet

The Internet of Things (IoT) has led to rapid growth in smart cities. However, IoT botnet-based attacks against smart city systems are becoming more prevalent. Detection methods for IoT botnet-based attacks have been the subject of extensive research, but the identification of early-stage behaviour of the IoT botnet prior to any attack remains a largely unexplored area that could prevent any attack before it is launched. Few studies have addressed the early stages of IoT botnet detection using monolithic deep learning algorithms that could require more time for training and detection. We, however, propose an edge-based deep learning system for the detection of the early stages of IoT botnets in smart cities. The proposed system, which we call EDIT (Edge-based Detection of early-stage IoT Botnet), aims to detect abnormalities in network communication traffic caused by early-stage IoT botnets based on the modular neural network (MNN) method at multi-access edge computing (MEC) servers. MNN can improve detection accuracy and efficiency by leveraging parallel computing on MEC. According to the findings, EDIT has a lower false-negative rate compared to a monolithic approach and other studies. At the MEC server, EDIT takes as little as 16 ms for the detection of an IoT botnet.

A Conformable Moments-Based Deep Learning System for Forged Handwriting Detection.

Detecting forged handwriting is important in a wide variety of machine learning applications, and it is challenging when the input images are degraded with noise and blur. This article presents a new model based on conformable moments (CMs) and deep ensemble neural networks (DENNs) for forged handwriting detection in noisy and blurry environments. Since CMs involve fractional calculus with the ability to model nonlinearities and geometrical moments as well as preserving spatial relationships between pixels, fine details in images are preserved. This motivates us to introduce a DENN classifier, which integrates stenographic kernels and spatial features to classify input images as normal (original, clean images), altered (handwriting changed through copy-paste and insertion operations), noisy (added noise to original image), blurred (added blur to original image), altered-noise (noise is added to the altered image), and altered-blurred (blur is added to the altered image). To evaluate our model, we use a newly introduced dataset, which comprises handwritten words altered at the character level, as well as several standard datasets, namely ACPR 2019, ICPR 2018-FDC, and the IMEI dataset. The first two of these datasets include handwriting samples that are altered at the character and word levels, and the third dataset comprises forged International Mobile Equipment Identity (IMEI) numbers. Experimental results demonstrate that the proposed method outperforms the existing methods in terms of classification rate.

Fully automated deep learning system for osteoporosis screening using chest computed tomography images.

Osteoporosis, a disease stemming from bone metabolism irregularities, affects approximately 200 million people worldwide. Timely detection of osteoporosis is pivotal in grappling with this public health challenge. Deep learning (DL), emerging as a promising methodology in the field of medical imaging, holds considerable potential for the assessment of bone mineral density (BMD). This study aimed to propose an automated DL framework for BMD assessment that integrates localization, segmentation, and ternary classification using various dominant convolutional neural networks (CNNs). In this retrospective study, a cohort of 2,274 patients underwent chest computed tomography (CT) was enrolled from January 2022 to June 2023 for the development of the integrated DL system. The study unfolded in 2 phases. Initially, 1,025 patients were selected based on specific criteria to develop an automated segmentation model, utilizing 2 VB-Net networks. Subsequently, a distinct cohort of 902 patients was employed for the development and testing of classification models for BMD assessment. Then, 3 distinct DL network architectures, specifically DenseNet, ResNet-18, and ResNet-50, were applied to formulate the 3-classification BMD assessment model. The performance of both phases was evaluated using an independent test set consisting of 347 individuals. Segmentation performance was evaluated using the Dice similarity coefficient; classification performance was appraised using the receiver operating characteristic (ROC) curve. Furthermore, metrics such as the area under the curve (AUC), accuracy, and precision were meticulously calculated. In the first stage, the automatic segmentation model demonstrated excellent segmentation performance, with mean Dice surpassing 0.93 in the independent test set. In the second stage, both the DenseNet and ResNet-18 demonstrated excellent diagnostic performance in detecting bone status. For osteoporosis, and osteopenia, the AUCs were as follows: DenseNet achieved 0.94 [95% confidence interval (CI): 0.91-0.97], and 0.91 (95% CI: 0.87-0.94), respectively; ResNet-18 attained 0.96 (95% CI: 0.92-0.98), and 0.91 (95% CI: 0.87-0.94), respectively. However, the ResNet-50 model exhibited suboptimal diagnostic performance for osteopenia, with an AUC value of only 0.76 (95% CI: 0.69-0.80). Alterations in tube voltage had a more pronounced impact on the performance of the DenseNet. In the independent test set with tube voltage at 100 kVp images, the accuracy and precision of DenseNet decreased on average by approximately 14.29% and 18.82%, respectively, whereas the accuracy and precision of ResNet-18 decreased by about 8.33% and 7.14%, respectively. The state-of-the-art DL framework model offers an effective and efficient approach for opportunistic osteoporosis screening using chest CT, without incurring additional costs or radiation exposure.

Applying social media in emergency response: an attention-based bidirectional deep learning system for location reference recognition in disaster tweets

Applying social media in emergency response: an attention-based bidirectional deep learning system for location reference recognition in disaster tweets

Automated detection of otosclerosis with interpretable deep learning using temporal bone computed tomography images

ObjectiveThis study aimed to develop an automated detection schema for otosclerosis with interpretable deep learning using temporal bone computed tomography images. MethodsWith approval from the institutional review board, we retrospectively analyzed high-resolution computed tomography scans of the temporal bone of 182 participants with otosclerosis (67 male subjects and 115 female subjects; average age, 36.42 years) and 157 participants without otosclerosis (52 male subjects and 102 female subjects; average age, 30.61 years) using deep learning. Transfer learning with the pretrained VGG19, Mask RCNN, and EfficientNet models was used. In addition, 3 clinical experts compared the system's performance by reading the same computed tomography images for a subset of 35 unseen subjects. An area under the receiver operating characteristic curve and a saliency map were used to further evaluate the diagnostic performance. ResultsIn prospective unseen test data, the diagnostic performance of the automatically interpretable otosclerosis detection system at the optimal threshold was 0.97 and 0.98 for sensitivity and specificity, respectively. In comparison with the clinical acumen of otolaryngologists at P < 0.05, the proposed system was not significantly different. Moreover, the area under the receiver operating characteristic curve for the proposed system was 0.99, indicating satisfactory diagnostic accuracy. ConclusionOur research develops and evaluates a deep learning system that detects otosclerosis at a level comparable with clinical otolaryngologists. Our system is an effective schema for the differential diagnosis of otosclerosis in computed tomography examinations.

Image-based remote evaluation of PASI scores with psoriasis by the YOLO-v4 algorithm.

As a chronic relapsing disease, psoriasis is characterized by widespread skin lesions. The Psoriasis Area and Severity Index (PASI) is the most frequently utilized tool for evaluating the severity of psoriasis in clinical practice. Nevertheless, long-term monitoring and precise evaluation pose difficulties for dermatologists and patients, which is time-consuming, subjective and prone to evaluation bias. To develop a deep learning system with high accuracy and speed to assist PASI evaluation, we collected 2657 high-quality images from 1486 psoriasis patients, and images were segmented and annotated. Then, we utilized the YOLO-v4 algorithm to establish the model via four modules, we also conducted a human-computer comparison through quadratic weighted Kappa (QWK) coefficients and intra-class correlation coefficients (ICC). The YOLO-v4 algorithm was selected for model training and optimization compared with the YOLOv3, RetinaNet, EfficientDet and Faster_rcnn. The model evaluation results of mean average precision (mAP) for various lesion features were as follows: erythema, mAP = 0.903; scale, mAP = 0.908; and induration, mAP = 0.882. In addition, the results of human-computer comparison also showed a median consistency for the skin lesion severity and an excellent consistency for the area and PASI score. Finally, an intelligent PASI app was established for remote disease assessment and course management, with a pleasurable agreement with dermatologists. Taken together, we proposed an intelligent PASI app based on the image YOLO-v4 algorithm that can assist dermatologists in long-term and objective PASI scoring, shedding light on similar clinical assessments that can be assisted by computers in a time-saving and objective manner.

Efficient breast cancer diagnosis using multi‐level progressive feature aggregation based deep transfer learning system

AbstractBreast cancer is a worldwide fatal disease that exists mostly among women. The deep learning technique has proven its effectiveness, but the performance of the existing deep learning systems is quite compromising. In this work, a deep transfer learning system is suggested for efficient breast cancer classification from histopathology images. This system is based on a novel multi‐level progressive feature aggregation (MPFA) and a spatial domain learning approach. The combination of a pretrained Resnet101 backbone network with MPFA is implemented to extract more significant features. In addition, a mixed‐dilated spatial domain learning network (MSLN) is further incorporated to enhance the receptive field and increase discrimination between features. The proposed method achieved superior performance as compared to the existing state‐of‐the‐art methods, offering 99.24% accuracy, a 98.79% F‐1 score, 98.59% precision, and 98.99% recall values over BreaKHis dataset. An ablation study is carried out over the ICIAR2018 dataset to verify the generalizability and effectiveness of the system.

Deep ocular tumor classification model using cuckoo search algorithm and Caputo fractional gradient descent

While digital ocular fundus images are commonly used for diagnosing ocular tumors, interpreting these images poses challenges due to their complexity and the subtle features specific to tumors. Automated detection of ocular tumors is crucial for timely diagnosis and effective treatment. This study investigates a robust deep learning system designed for classifying ocular tumors. The article introduces a novel optimizer that integrates the Caputo fractional gradient descent (CFGD) method with the cuckoo search algorithm (CSA) to enhance accuracy and convergence speed, seeking optimal solutions. The proposed optimizer’s performance is assessed by training well-known Vgg16, AlexNet, and GoogLeNet models on 400 fundus images, equally divided between benign and malignant classes. Results demonstrate the significant potential of the proposed optimizer in improving classification accuracy and convergence speed. In particular, the mean accuracy attained by the proposed optimizer is 86.43%, 87.42%, and 87.62% for the Vgg16, AlexNet, and GoogLeNet models, respectively. The performance of our optimizer is compared with existing approaches, namely stochastic gradient descent with momentum (SGDM), adaptive momentum estimation (ADAM), the original cuckoo search algorithm (CSA), Caputo fractional gradient descent (CFGD), beetle antenna search with ADAM (BASADAM), and CSA with ADAM (CSA-ADAM). Evaluation criteria encompass accuracy, robustness, consistency, and convergence speed. Comparative results highlight significant enhancements across all metrics, showcasing the potential of deep learning techniques with the proposed optimizer for accurately identifying ocular tumors. This research contributes significantly to the development of computer-aided diagnosis systems for ocular tumors, emphasizing the benefits of the proposed optimizer in medical image classification domains.

Automatic Identification of Pangolin Behavior Using Deep Learning Based on Temporal Relative Attention Mechanism.

With declining populations in the wild, captive rescue and breeding have become one of the most important ways to protect pangolins from extinction. At present, the success rate of artificial breeding is low, due to the insufficient understanding of the breeding behavior characteristics of pangolins. The automatic recognition method based on machine vision not only monitors for 24 h but also reduces the stress response of pangolins. This paper aimed to establish a temporal relation and attention mechanism network (Pangolin breeding attention and transfer network, PBATn) to monitor and recognize pangolin behaviors, including breeding and daily behavior. There were 11,476 videos including breeding behavior and daily behavior that were divided into training, validation, and test sets. For the training set and validation set, the PBATn network model had an accuracy of 98.95% and 96.11%, and a loss function value of 0.1531 and 0.1852. The model is suitable for a 2.40 m × 2.20 m (length × width) pangolin cage area, with a nest box measuring 40 cm × 30 cm × 30 cm (length × width × height) positioned either on the left or right side inside the cage. A spherical night-vision monitoring camera was installed on the cage wall at a height of 2.50 m above the ground. For the test set, the mean Average Precision (mAP), average accuracy, average recall, average specificity, and average F1 score were found to be higher than SlowFast, X3D, TANet, TSN, etc., with values of 97.50%, 99.17%, 97.55%, 99.53%, and 97.48%, respectively. The recognition accuracies of PBATn were 94.00% and 98.50% for the chasing and mounting breeding behaviors, respectively. The results showed that PBATn outperformed the baseline methods in all aspects. This study shows that the deep learning system can accurately observe pangolin breeding behavior and it will be useful for analyzing the behavior of these animals.

Reduction of false positives using zone-specific prostate-specific antigen density for prostate MRI-based biopsy decision strategies

ObjectivesTo develop and test zone-specific prostate-specific antigen density (sPSAD) combined with PI-RADS to guide prostate biopsy decision strategies (BDS).MethodsThis retrospective study included consecutive patients, who underwent prostate MRI and biopsy (01/2012–10/2018). The whole gland and transition zone (TZ) were segmented at MRI using a retrained deep learning system (DLS; nnU-Net) to calculate PSAD and sPSAD, respectively. Additionally, sPSAD and PI-RADS were combined in a BDS, and diagnostic performances to detect Grade Group ≥ 2 (GG ≥ 2) prostate cancer were compared. Patient-based cancer detection using sPSAD was assessed by bootstrapping with 1000 repetitions and reported as area under the curve (AUC). Clinical utility of the BDS was tested in the hold-out test set using decision curve analysis. Statistics included nonparametric DeLong test for AUCs and Fisher-Yates test for remaining performance metrics.ResultsA total of 1604 patients aged 67 (interquartile range, 61–73) with 48% GG ≥ 2 prevalence (774/1604) were evaluated. By employing DLS-based prostate and TZ volumes (DICE coefficients of 0.89 (95% confidence interval, 0.80–0.97) and 0.84 (0.70–0.99)), GG ≥ 2 detection using PSAD was inferior to sPSAD (AUC, 0.71 (0.68–0.74)/0.73 (0.70–0.76); p < 0.001). Combining PI-RADS with sPSAD, GG ≥ 2 detection specificity doubled from 18% (10–20%) to 43% (30–44%; p < 0.001) with similar sensitivity (93% (89–96%)/97% (94–99%); p = 0.052), when biopsies were taken in PI-RADS 4-5 and 3 only if sPSAD was ≥ 0.42 ng/mL/cc as compared to all PI-RADS 3-5 cases. Additionally, using the sPSAD-based BDS, false positives were reduced by 25% (123 (104–142)/165 (146–185); p < 0.001).ConclusionUsing sPSAD to guide biopsy decisions in PI-RADS 3 lesions can reduce false positives at MRI while maintaining high sensitivity for GG ≥ 2 cancers.Clinical relevance statementTransition zone-specific prostate-specific antigen density can improve the accuracy of prostate cancer detection compared to MRI assessments alone, by lowering false-positive cases without significantly missing men with ISUP GG ≥ 2 cancers.Key Points• Prostate biopsy decision strategies using PI-RADS at MRI are limited by a substantial proportion of false positives, not yielding grade group ≥ 2 prostate cancer.• PI-RADS combined with transition zone (TZ)-specific prostate-specific antigen density (PSAD) decreased the number of unproductive biopsies by 25% compared to PI-RADS only.• TZ-specific PSAD also improved the specificity of MRI-directed biopsies by 9% compared to the whole gland PSAD, while showing identical sensitivity.

PointCVaR: Risk-Optimized Outlier Removal for Robust 3D Point Cloud Classification

With the growth of 3D sensing technology, the deep learning system for 3D point clouds has become increasingly important, especially in applications such as autonomous vehicles where safety is a primary concern. However, there are growing concerns about the reliability of these systems when they encounter noisy point clouds, either occurring naturally or introduced with malicious intent. This paper highlights the challenges of point cloud classification posed by various forms of noise, from simple background noise to malicious adversarial/backdoor attacks that can intentionally skew model predictions. While there's an urgent need for optimized point cloud denoising, current point outlier removal approaches, an essential step for denoising, rely heavily on handcrafted strategies and are not adapted for higher-level tasks, such as classification. To address this issue, we introduce an innovative point outlier cleansing method that harnesses the power of downstream classification models. Using gradient-based attribution analysis, we define a novel concept: point risk. Drawing inspiration from tail risk minimization in finance, we recast the outlier removal process as an optimization problem, named PointCVaR. Extensive experiments show that our proposed technique not only robustly filters diverse point cloud outliers but also consistently and significantly enhances existing robust methods for point cloud classification. A notable feature of our approach is its effectiveness in defending against the latest threat of backdoor attacks in point clouds.

Design Analysis of Stroke Risk Prediction Model Employing the Hybrid Structure Implementation of Deep Transfer Learning System

Design Analysis of Stroke Risk Prediction Model Employing the Hybrid Structure Implementation of Deep Transfer Learning System

A novel approach for self-driving car in partially observable environment using life long reinforcement learning

Despite ground-breaking advancements in robotics, gaming, and other challenging domains, reinforcement learning still faces significant challenges in solving dynamic, open-world problems. Since reinforcement learning algorithms usually perform poorly when exposed to new tasks outside of their data distribution, continuous learning algorithms have drawn significant attention. In parallel with work on lifelong learning algorithms, there is a need for challenging environments, properly planned trials, and metrics to measure research success. In this context, a Deep Asynchronous Autonomous Learning System (DAALS) is proposed in this paper for training a self-driving car in a partially observable environment for real-world scenarios in a continuous state-action space. To cater to three different use cases, three different algorithms were used. To train their agents for learning and upgrading discrete state policies, DAALS used the Asynchronous Advantage Stager Reviewer (AASR) algorithm. To train its agent for continuous state spaces, DAALS also uses an Extensive Deterministic Policy Gradient (EDPG) algorithm. To train the agent in a lifelong form of learning for partially observable environments, DAALS uses a Deep Deterministic Policy Gradient Novel Lifelong Learning Algorithm (DDPGNLLA). The system offers flexibility to the user to train the agents for both discrete and continuous state-action spaces. Compared to previous models in continuous state-action spaces, Deep deterministic policy gradient lifelong learning algorithm outperforms previous models by 46.09%. Furthermore, the Deep Asynchronous Autonomous System tends to outperform all previous reinforcement learning algorithms, making our proposed approach a real-world solution. As DAALS has tested on number of different environments it provides the insights on how modern Artificial Intelligence (AI) solutions can be generalized making it one of the better solutions for AI general domain problems.

Deep Learning for Detection of Proper Utilization and Adequacy of Personal Protective Equipment in Manufacturing Teaching Laboratories

Occupational sectors are perennially challenged by the potential for workplace accidents, particularly in roles involving tools and machinery. A notable cause of such accidents is the inadequate use of Personal Protective Equipment (PPE), essential in preventing injuries and illnesses. This risk is not confined to workplaces alone but extends to educational settings with practical activities, like manufacturing teaching laboratories in universities. Current methods for monitoring and ensuring proper PPE usage especially in the laboratories are limited, lacking in real-time and accurate detection capabilities. This study addresses this gap by developing a visual-based, deep learning system specifically tailored for assessing PPE usage in manufacturing teaching laboratories. The method of choice for object detection in this study is You Only Look Once (YOLO) algorithms, encompassing YOLOv4, YOLOv5, and YOLOv6. YOLO processes images in a single pass through its architecture, in which its efficiency allows for real-time detection. The novel contribution of this study lies in its computer vision models, adept at not only detecting compliance but also assessing adequacy of PPE usage. The result indicates that the proposed computer vision models achieve high accuracy for detection of PPE usage compliance and adequacy with a mAP value of 0.757 and an F1-score of 0.744, obtained with the YOLOv5 model. The implementation of a deep learning system for PPE compliance in manufacturing teaching laboratories could markedly improve safety, preventing accidents and injuries through real-time compliance monitoring. Its effectiveness and adaptability could set a precedent for safety protocols in various educational settings, fostering a wider culture of safety and compliance.

Brandt's vole hole detection and counting method based on deep learning and unmanned aircraft system.

Rodents are essential to the balance of the grassland ecosystem, but their population outbreak can cause major economic and ecological damage. Rodent monitoring is crucial for its scientific management, but traditional methods heavily depend on manual labor and are difficult to be carried out on a large scale. In this study, we used UAS to collect high-resolution RGB images of steppes in Inner Mongolia, China in the spring, and used various object detection algorithms to identify the holes of Brandt's vole (Lasiopodomys brandtii). Optimizing the model by adjusting evaluation metrics, specifically, replacing classification strategy metrics such as precision, recall, and F1 score with regression strategy-related metrics FPPI, MR, and MAPE to determine the optimal threshold parameters for IOU and confidence. Then, we mapped the distribution of vole holes in the study area using position data derived from the optimized model. Results showed that the best resolution of UAS acquisition was 0.4cm pixel-1, and the improved labeling method improved the detection accuracy of the model. The FCOS model had the highest comprehensive evaluation, and an R2 of 0.9106, RMSE of 5.5909, and MAPE of 8.27%. The final accuracy of vole hole counting in the stitched orthophoto was 90.20%. Our work has demonstrated that UAS was able to accurately estimate the population of grassland rodents at an appropriate resolution. Given that the population distribution we focus on is important for a wide variety of species, our work illustrates a general remote sensing approach for mapping and monitoring rodent damage across broad landscapes for studies of grassland ecological balance, vegetation conservation, and land management.

An interpretable machine learning system for colorectal cancer diagnosis from pathology slides

Considering the profound transformation affecting pathology practice, we aimed to develop a scalable artificial intelligence (AI) system to diagnose colorectal cancer from whole-slide images (WSI). For this, we propose a deep learning (DL) system that learns from weak labels, a sampling strategy that reduces the number of training samples by a factor of six without compromising performance, an approach to leverage a small subset of fully annotated samples, and a prototype with explainable predictions, active learning features and parallelisation. Noting some problems in the literature, this study is conducted with one of the largest WSI colorectal samples dataset with approximately 10,500 WSIs. Of these samples, 900 are testing samples. Furthermore, the robustness of the proposed method is assessed with two additional external datasets (TCGA and PAIP) and a dataset of samples collected directly from the proposed prototype. Our proposed method predicts, for the patch-based tiles, a class based on the severity of the dysplasia and uses that information to classify the whole slide. It is trained with an interpretable mixed-supervision scheme to leverage the domain knowledge introduced by pathologists through spatial annotations. The mixed-supervision scheme allowed for an intelligent sampling strategy effectively evaluated in several different scenarios without compromising the performance. On the internal dataset, the method shows an accuracy of 93.44% and a sensitivity between positive (low-grade and high-grade dysplasia) and non-neoplastic samples of 0.996. On the external test samples varied with TCGA being the most challenging dataset with an overall accuracy of 84.91% and a sensitivity of 0.996.

Development and validation of a deep learning system for detection of small bowel pathologies in capsule endoscopy: a pilot study in a Singapore institution.

Deep learning models can assess the quality of images and discriminate among abnormalities in small bowel capsule endoscopy (CE), reducing fatigue and the time needed for diagnosis. They serve as a decision support system, partially automating the diagnosis process by providing probability predictions for abnormalities. We demonstrated the use of deep learning models in CE image analysis, specifically by piloting a bowel preparation model (BPM) and an abnormality detection model (ADM) to determine frame-level view quality and the presence of abnormal findings, respectively. We used convolutional neural network-based models pretrained on large-scale open-domain data to extract spatial features of CE images that were then used in a dense feed-forward neural network classifier. We then combined the open-source Kvasir-Capsule dataset (n = 43) and locally collected CE data (n = 29). Model performance was compared using averaged five-fold and two-fold cross-validation for BPMs and ADMs, respectively. The best BPM model based on a pre-trained ResNet50 architecture had an area under the receiver operating characteristic and precision-recall curves of 0.969±0.008 and 0.843±0.041, respectively. The best ADM model, also based on ResNet50, had top-1 and top-2 accuracies of 84.03±0.051 and 94.78±0.028, respectively. The models could process approximately 200-250 images per second and showed good discrimination on time-critical abnormalities such as bleeding. Our pilot models showed the potential to improve time to diagnosis in CE workflows. To our knowledge, our approach is unique to the Singapore context. The value of our work can be further evaluated in a pragmatic manner that is sensitive to existing clinician workflow and resource constraints.