Scopus Indexed Publications

Paper Details


Title
ProtoViT-Net: A Convolution–Transformer Prototypical Network for Few-Shot Multicancer Image Classification

Author
Md. Darun Nayeem,

Email

Abstract

Cancer remains a major global health burden, where early and accurate diagnosis proves to be lifesaving. Despite recent progress in deep learning based histopathological analysis, existing research faces two persistent limitations: (i) dependence on large, expert-annotated datasets that are scarce in clinical practice, and (ii) lack of real-time, resource-efficient deployment suitable for practical diagnostic environments. To address these challenges, this study introduces ProtoViT-Net, a lightweight convolution and transformer prototypical framework for few-shot multicancer classification. The model employs a hybrid encoder that combines depth-wise separable convolutional layers to capture nuclei-level morphological textures with mechanisms that jointly learn local and global feature dependencies through efficient self-attention. These complementary representations are embedded into a prototypical metric space, enabling distance-based classification under an episodic metalearning paradigm. The proposed model is evaluated across diverse histopathological datasets including lung, colon, breast, and skin cancers, and demonstrates 99.42% classification accuracy under a 3-way, 10-shot episodic few-shot evaluation protocol, while maintaining superior computational efficiency compared with representative baseline methods. With fewer than 37 K parameters, ProtoViT-Net demonstrates the feasibility of real-time web and mobile deployment through a lightweight prototype implementation. Comprehensive deployment benchmarking, including cross-device TensorFlow Lite optimization, runtime memory profiling, and large-scale mobile performance evaluation, is reserved for future work. By unifying morphological representation learning with metric-based adaptation, ProtoViT-Net establishes a data-efficient and clinically deployable pathway toward intelligent digital pathology.


Keywords

Journal or Conference Name
Journal of Engineering (United Kingdom)

Publication Year
2026

Indexing
scopus