Explainable Lightweight Model Using Low-Rank and Convolutional Block Attention for Pancreatic Cancer Diagnosis

Sep 22, 2026

Upload by Angel

Early and accurate pancreatic cancer (PC) detection remains a major clinical challenge, largely due to subtle early-stage CT appearances, low contrast with surrounding tissue, and variability in scanning protocols. The study proposes a hybrid, lightweight deep learning framework for automated CT image classification, integrating MobileNetV3Small, convolutional block attention, and Low-rank Attention with Shared Efficient Representations (LASER), plus transformer encoding and cross-type interaction fusion. Evaluation on 18,942 CT images yields 99.34% accuracy, AUC-ROC 0.9996, Cohen's Kappa 0.9897, and MCC 0.9859, with explainability via Grad-CAM variants indicating clinically relevant focus regions.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Abstract

1 Introduction

1.1 Problem Introduction

The International Journal of Medical Robotics and Computer Assisted Surgery - ORIGINAL ARTICLE OPEN ACCESS Explainable Lightweight Model Using Low‐Rank and Convolutional Block Attention for Pancreatic Cancer Diagnosis Vishesh Tanwar1 | Bhisham Sharma1 | Dhirendra Prasad Yadav2 | Panos Liatsis3 1Chitkara University Institute of Engineering and Technology, Chitkara University, Rajpura, Punjab, India | 2Department of Computer Engineering & Applications, G.L.A. University, Mathura, Uttar Pradesh, India | 3Department of Computer Science, Center for Cyber Physical Systems, Khalifa University, Abu Dhabi, UAE Correspondence: Panos Liatsis ([panos.liatsis@ku.ac.ae](panos.liatsis@ku.ac.ae)) Received: 10 November 2025 | Revised: 24 March 2026 | Accepted: 1 April 2026 Keywords: attention | classification | deep learning | pancreatic cancer | transformer ABSTRACT 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. 1 | Introduction 1.1 | Problem Introduction Pancreatic cancer, particularly pancreatic ductal adenocarcinoma (PDAC), is one of the most lethal malignancies worldwide, characterised by late diagnosis, aggressive progression and limited treatment options, with 5‐year survival rates reported as low as 7% [1] . Despite advances in medical imaging, early detection remains a significant clinical challenge due to the subtle and non‐specific appearance of early‐stage tumours on computed tomography (CT) scans. The dismal prognosis is primarily attributed to late‐stage diagnosis, aggressive tumour progression, and the lack of effective early screening strategies. Although computed tomography (CT) plays a central role in the imaging and diagnostic assessment of pancreatic cancer, early detection remains a formidable clinical challenge. This is due to the subtle and non‐specific nature of early‐stage lesions, low contrast between normal and pathological tissues, and variability in CT acquisition protocols. Even highly experienced radiologists often struggle to distinguish malignant tissue from benign conditions or inflammatory changes, leading to potential misdiagnosis and delayed treatment [2] . Deep learning has emerged as a transformative technology in medical image analysis in recent years, particularly for tasks requiring high‐level feature representation and classification [3] . Unlike traditional ML approaches that rely on handcrafted This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made. © 2026 The Author(s) . The International Journal of Medical Robotics and Computer Assisted Surgery published by John Wiley & Sons Ltd. The International Journal of Medical Robotics and Computer Ass

FAQs

What clinical problem does the paper address?

It addresses the difficulty of early and accurate pancreatic cancer detection using computed tomography (CT), where early lesions are subtle and non-specific.

How is the proposed model structured?

It uses a lightweight MobileNetV3Small backbone with convolutional block attention and Low-rank Attention with Shared Efficient Representations (LASER), projects feature maps into token sequences for a transformer encoder, and fuses aggregated features via a cross-type interaction module.

What performance results does the model achieve and how is explainability handled?

On 18,942 CT images, it reaches 99.34% accuracy with AUC-ROC 0.9996, Cohen's Kappa 0.9897, and MCC 0.9859. Explainability uses Grad-CAM, Grad-CAM++, and attention visualization to indicate focus on clinically relevant regions.

Sep 22, 2026

Upload by Angel