Automated Segmentation of Bleeding Regions in Wireless Capsule Endoscopy Using CNN–Transformer Deep Learning Architectures

Transformer-Based Deep Learning Framework for Automated Lesion Detection in Capsule Endoscopy: A Comparative Study With CNN Architectures.

GOALS To compare a vision transformer with 2 convolutional neural network architectures for multiclass lesion classification in capsule endoscopy images. BACKGROUND Manual review of capsule endoscopy is time-consuming and subject to interobserver variability. Deep learning can automate lesion recognition; however, most prior capsule endoscopy work evaluates a small number of classes, and systematic comparisons between transformer and convolutional architectures across many lesion categories are limited. STUDY Two publicly available data sets (SEE-AI and Kvasir-Capsule) were merged and preprocessed to create a 21-class image data set (∼58,000 frames). Images were resized to 224 × 224 pixels and split using stratified sampling into training (n=40,587), validation (n=8696), and test (n=8696) sets. A pretrained Vision Transformer, DenseNet121, and ResNet50 were fine-tuned using categorical cross-entropy loss and Adam optimization, with early stopping. Performance was assessed using accuracy, macroaveraged precision, recall, F1 Score, and the area under the receiver operating characteristic curve. RESULTS On the independent test set, the vision transformer achieved 92.2% accuracy with macroaveraged precision/recall/F1-score of 0.92 and an area under the receiver operating characteristic curve of 0.99. DenseNet121 achieved 74.0% accuracy (F1-score 0.78; area under the receiver operating characteristic curve 0.85). ResNet50 achieved 38.0% accuracy (F1-score 0.40; area under the receiver operating characteristic curve 0.55). CONCLUSIONS In this merged 21-class capsule endoscopy image data set, the vision transformer achieved higher frame-level classification performance than DenseNet121 and ResNet50 under the present experimental conditions. Importantly, the data set was split at the image level rather than at the patient or procedure level, because frames from the same examination may be correlated; therefore, performance figures likely reflect benchmark results on this frame-level public data set and should not be interpreted as estimates of patient-level generalization or as evidence of definitive architectural superiority. These findings support further evaluation of transformer-based approaches, but grouped reanalysis, external validation, and workflow-oriented studies are required before clinical implementation.

An End-to-End Deep Learning System for Gastrointestinal Bleeding Detection and Quantification in Wireless Capsule Endoscopy

Background/Objectives: Gastrointestinal bleeding is a critical finding in wireless capsule endoscopy (WCE), but manual examination of thousands of image frames is labor-intensive, time-consuming, and susceptible to missed lesions. This study aimed to develop and evaluate a comprehensive deep-learning framework for automated bleeding detection, localization, and quantitative assessment in WCE images. Methods: The proposed framework integrates three complementary deep-learning models: (i) a custom two-dimensional convolutional neural network (2D-CNN) for frame-level bleeding classification, (ii) a three-dimensional convolutional neural network (3D-CNN) for sequence-level analysis by exploiting temporal information from consecutive frames, and (iii) a U-Net architecture for pixel-level segmentation and bleeding-area quantification. The models were trained and evaluated using expert-annotated WCE datasets with pixel-level ground-truth masks. Results: The proposed 2D-CNN and 3D-CNN achieved excellent classification performance, with areas under the receiver operating characteristic curve (AUCs) of 0.9986 and 0.9971, respectively. The U-Net model achieved a Dice similarity coefficient of 0.93, an intersection-over-union (IoU) of 0.8677, and an overall segmentation accuracy of 97.25%. The integrated framework outperformed previously reported methods, demonstrating robust performance for bleeding detection, localization, and quantitative assessment. Conclusions: The proposed end-to-end deep-learning framework enables accurate automated bleeding detection, localization, and severity quantification in WCE images. By reducing the burden of manual image review, improving diagnostic consistency, and providing objective bleeding assessment, the framework has strong potential to support clinical decision-making and enhance gastrointestinal diagnostic workflows.

Intelligent recognition and segmentation of anatomical structures in spinal endoscopy: a deep learning approach with 1000 annotated images.

A novel medical image segmentation network for colorectal polyp small targets and fuzzy boundaries

Abstract Objectives To address the challenges of complex feature variations and unclear boundary definitions between segmented targets and surrounding regions in medical images, a novel segmentation model based on Deformable Large Kernel Convolutional Attention (D-LKA) and Transformer is proposed. Methods The model first uses Vision Transformer as the encoder to enhance the ability to capture global information, overcoming the limitations of convolutional neural networks’ receptive field. In the decoder, a D-LKA decoder with deformable large kernel convolution attention is used, allowing the model to adapt to complex target features. Finally, the TRR module is introduced to coordinate information transfer between the convolutional neural network and Transformer, reducing semantic loss. Results The model is trained, validated, and tested on the Kvasir-SEG colon polyp dataset, with multiple ablation experiments. To validate generalization, experiments are also conducted on the CVC-ClinicDB dataset. Experimental results show that TDU-Net outperforms other methods in both segmentation accuracy and generalization. Conclusions TDU-Net achieves excellent segmentation results, addressing small target and incomplete feature extraction issues, significantly improving clinical diagnosis efficiency and accuracy.

Vision Transformer Outperforms CNN Architectures in Binary Skin Lesion Classification: An Eight-Model Controlled Deep Learning Study

Skin cancer is one of the life-threatening malignancies in the world, timely detection of which is directly proportional to the survival of patients. Traditional dermoscopic diagnosis suffers inter-observer variability, lack of specialists and is not scalable in resource limited healthcare environments. The end-to-end hierarchical feature learning provided by deep learning is a transformative solution to the learned dermoscopic image corpora. The paper empirically comparatively examines eight binary skin lesion classifiers benign versus malignant of a dataset of 2,637 training images, and 661 held out test images. The assessed architectures are located on a wide design range a custom CNN trained using fresh data, two ResNet18 transfer learning pipelines, DenseNet121, MobileNetV3-Large, ViT-Small/16 (ImageNet-21K), ConvNeXt-Tiny (ImageNet-12K) and EfficientNetV2-S (ImageNet-21K). All the models are trained with the same conditions involving stratified splitting, Weighted Random Sampler, two-stage fine tuning with discriminative learning rates, Automatic Mixed Precision, and early stopping. It is evaluated using six metrics accuracy, per-class precision, recall, F1-score, ROC-AUC, and PR-AUC. The highest test accuracy (91.53) and macro-F1 (0.907) is attained with EfficientNetV2-S. ViT-Small/16 has the best ROC-AUC (0.9723) and PR-AUC (0.9699), which proves the effectiveness of Vision Transformer in threshold-free probabilistic discrimination the clinically decisive measure in screening applications. The three contemporary timm-based models have consistently reached ROC-AUC 0.95 and above, but the legacy CNN models are at 0.56 even though the legacy CNN models are at competitive accuracy of 89-90%. MobileNetV3-Large yields a false negative rate of 61 (20% miss rate), which highlights the clinical risk of aggressive model compression. The findings simplify the selection of the model in clinical studies in implementing dermoscopy, suggesting that ViT-Small/16 should be used as a probabilistic-ranked malignancy screening model, and EfficientNetV2-S should be used as a fixed threshold binary triage model in a telemedicine system.

A Lightweight CNN Framework for Real Time Liver Lesion Classification from CT Images with YOLOv8 Guided Localization

In order to improve patient outcomes, rapid and repeatable characterisation of hepatic tumours by computed tomography (CT) is essential. Liver cancer is still one of the most deadly cancers in the world. Although traditional volumetric (3D) segmented networks accurately identify lesions, their utility for real time screenings on commodity hardware is restricted due to their high computational cost and delay. A compact a convolutional neural network (CNN) which classifies hepatic regions as benign, malignant, or normal is the main contribution of this works lightweight classification framework. An upstream, anchor less YOLOv8 stage provides candidate regions for lesion localisation. A semi supervised U-Net pipeline is used to transform volumetric research from the LiTS, 3DIRCADb, along with CHAOS collections into annotated 2D axial slices. While the upstream localisation step runs at about 30 ms per slice, the CNN classification stage achieves a mean per class accuracy of about 95% (0.96 normal, 0.94 benign, and 0.95 malignant) with minor, symmetric inter class confusion on the curated test set. A Flask web interface that overlays areas, confidence scores, as well as lesion labels in real time, provides access to the framework. The findings show that early liver cancer screening is a good fit for a lightweight, scalable classification pipeline, future research should focus on improving small lesion localisation.