Integrating expert guidance with gradual moment approximation (GMAp)-enhanced transfer learning for improved pancreatic cancer classification

Despite significant research efforts, pancreatic cancer remains a formidable foe. To address the critical need for improved diagnostics, this study presents a novel approach that integrates expert guidance with computer-aided imaging for fine needle aspiration (FNA). A meticulously curated computed tomography (CT) dataset of ground truth images, focusing on key subregions of the pancreas, was established in collaboration with medical professionals. The images provided the training ground for a novel diagnostic model equipped with the gradual moment approximation (GMAp) optimization algorithm, designed to enhance the precision of cancer detection. By efficiently transferring knowledge from pre-trained models, the proposed model achieved remarkable accuracy (98.16%) in classifying CT images across distinct cancerous pancreatic subregions (head, body, and tail) and healthy pancreas. Extensive evaluations against diverse pre-trained models and benchmark medical databases: medical segmentation decathlon, clinical proteomic tumor analysis consortium pancreatic ductal adenocarcinoma, and pancreas-computed tomography proved the model's robustness and superior F1-scores compared to existing approaches. The experiment demonstrates that the deep learning-based 4-class classification outperforms state-of-the-art machine learning-based method by 3.66% in terms of accuracy. This efficiency, coupled with rigorous testing, paves the way for seamless integration into clinical workflows, potentially enabling earlier and more accurate pancreatic cancer diagnoses.