Vision Transformers (ViTs) are one of the powerful tools in medical imaging, providing new possibilities for pancreatic cancer diagnosis. In recent years, several studies have reported deep learning (DL) techniques to computed tomography (CT) images for pancreatic cancer diagnosis using ViT-based architectures. Existing methods often suffer from high computational complexity and there are limitations in reducing false negatives, particularly for malignant lesion. This paper proposes an intelligent pancreatic tumor detection framework called Rotary Positional Siamese Vision Transformer (RPSViT), designed to accurately detect and classify pancreatic tumors by effectively localizing abnormalities in CT scan images. RPSViT employs a patch-based approach, dividing input images into fixed-size patches that are treated as tokens via linear patch embedding. Rotary positional embedding is then incorporated to capture better spatial relationships within the images, thereby enhancing tumor localisation accuracy. The Siamese Transformer Encoder extracts high-level feature vectors from the input samples and performs disease classification. The model was trained and evaluated on Pancreatic-CT scan images from The Cancer Imaging Archive (TCIA) datasets and Medical Segmentation Decathlon (MSD) datasets using a 5-fold cross-validation. Experimental results shows that the proposed RPSViT achieves a mean accuracy of 96.97 ± 1.81%, sensitivity of 96.06 ± 2.38%, specificity of 100.00 ± 0.00%, and mean AUC of 0.9989 ± 0.0015. Additionally, the framework attains an F1-score of 0.9798 ± 0.0125, Matthews correlation coefficient (MCC) of 0.9225 ± 0.0414, Cohen's kappa coefficient of 0.9188 ± 0.0448, average precision of 0.9997 ± 0.0004, and Jaccard index of 0.9606 ± 0.0238. These RPSViT performance results shows that it effectively bridges advanced transformer architectures with practical medical diagnostics, providing more accurate and automated tools in pancreatic oncology.
Abinaya et al. (2026) studied this question.