The international journal of medical robotics + computer assisted surgery : MRCAS · 2026
Explainable Lightweight Model Using Low-Rank and Convolutional Block Attention for Pancreatic Cancer Diagnosis.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Tanwar Vishesh, Sharma Bhisham, Yadav Dhirendra Prasad, Liatsis Panos
Аннотация
BACKGROUND: Early and accurate pancreatic cancer (PC) detection remains a major clinical challenge.
METHODS: We introduce a novel hybrid deep learning framework for automated classification of CT images, which requires fewer computational resources while achieving high diagnostic performance. We integrated a lightweight MobileNetV3Small backbone with a convolutional block attention module and Low-rank Attention with Shared Efficient Representations (LASER) to enhance feature representation. Feature maps are projected via a 1 × 1 convolution into token sequences and processed through a transformer encoder to capture long-range dependencies. A parallel global average pooling extracts aggregated features, fused using a cross-type interaction (CTI) module.
RESULTS: The model was evaluated on 18,942 CT images and achieved 99.34% accuracy, AUC-ROC of 0.9996, Cohen's Kappa of 0.9897, and MCC of 0.9859, outperforming ResNet50, EfficientNetB0, and ViT variants with only 1.26 million parameters.
CONCLUSIONS: Explainability analyses using Grad-CAM, Grad-CAM++, and attention visualisation suggest that the model focuses on clinically relevant regions.
METHODS: We introduce a novel hybrid deep learning framework for automated classification of CT images, which requires fewer computational resources while achieving high diagnostic performance. We integrated a lightweight MobileNetV3Small backbone with a convolutional block attention module and Low-rank Attention with Shared Efficient Representations (LASER) to enhance feature representation. Feature maps are projected via a 1 × 1 convolution into token sequences and processed through a transformer encoder to capture long-range dependencies. A parallel global average pooling extracts aggregated features, fused using a cross-type interaction (CTI) module.
RESULTS: The model was evaluated on 18,942 CT images and achieved 99.34% accuracy, AUC-ROC of 0.9996, Cohen's Kappa of 0.9897, and MCC of 0.9859, outperforming ResNet50, EfficientNetB0, and ViT variants with only 1.26 million parameters.
CONCLUSIONS: Explainability analyses using Grad-CAM, Grad-CAM++, and attention visualisation suggest that the model focuses on clinically relevant regions.