Author ORCID Identifier

0009-0005-5650-9759

Author Linked-In Account

www.linkedin.com/in/pooja-g-ab771a306

Biosketch

Dr. Pooja G is a researcher in the field of Artificial Intelligence and Medical Image Analysis. She completed her doctoral research at the School of Computing, SASTRA Deemed University, Thanjavur, India. Her research interests include deep learning, computer vision, explainable artificial intelligence, graph and hypergraph neural networks, biomedical image analysis, and cancer diagnosis.

Her doctoral research focuses on the development of hypergraph-based deep neural frameworks for precise cancer subtyping and meta-visualization using histopathological and cytopathological images. She has contributed to several research works involving lung cancer, colon cancer, cervical cancer, blood cancer, medicinal plant identification, and plant disease detection using advanced machine learning and deep learning techniques. She is currently serving as an Assistant Professor in the School of Computing at SASTRA Deemed University, where she is actively involved in teaching, research, and academic activities.

Date of Award

3-3-2026

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.N.Sasikaladevi

Keywords

Deep Learning, Cancer, Subtyping, Histopathological Images, Depthwise Separable Convolutional Neural Network, Hypergraph Convolutional Neural Network, Explainable AI

Abstract

Accurate Cancer Subtyping is a cornerstone of modern oncology essential for effective diagnosis and guiding personalized treatment. Histopathological Images (HIs) which capture the microscopic structure of tissues are widely used for cancer detection and subtyping. Even though deep learning has made significant advances, existing HI based subtyping methods often focus on specific cancer types, lacking a generic framework.

A unified framework that can classify multiple cancers with high specificity is desperately needed. In response to these limitations, this thesis proposes a robust multi-cancer, multi-class subtyping framework called DSHGNet (Depthwise Separable Hypergraph Convolutional Neural Network) which integrates Depthwise Separable Convolutional Neural Networks (DSCNN) for deep feature extraction and k-order Feature Fusion Hypergraph Convolutional Neural Networks (kFF-HCNN) for precise classification.

DSCNN effectively extracts multi-scale features from histopathological images and kFF-HCNN enables accurate multi-class subtyping of lung, cervical, and colon cancers by modeling higher-order dependencies among features using k-order feature fusion strategy. To enhance interpretability and clinical transparency meta-visualization techniques such as Grad-CAM, occlusion sensitivity, and gradient attribution were employed for identifying class-discriminative regions.

The proposed framework's robustness and generalizability across different cancer types were demonstrated by a thorough evaluation on three benchmark histopathological datasets: Kather-2016 (colon), SIPaKMeD (cervical), and LC25000 (lung). Extensive experiments demonstrate that the proposed DSHGNet achieved an accuracy of 99.98% with precision, recall, F1-score, and specificity reaching 99.89%, 99.89%, 99.94%, and 99.94% respectively on LC25000 dataset.

For the SIPaKMeD dataset, the model yielded 99.31% accuracy, 99.38% recall, 98.97% precision, 99.34% F1-score, and 99.81% specificity. On the Kather-2016 dataset, DSHGNet attained 99.48% accuracy, 97.95% recall, 97.97% precision, 97.96% F1-score, and 99.76% specificity. These results demonstrate the generalizability and robustness of the proposed DSHGNet. The model consistently achieved specificity exceeding 99% thereby satisfies the most important clinical necessity and establishing itself as an ultra-specificity model. This research addresses critical gaps in computational pathology by offering a generic, ultra-specific, and interpretable framework for multi-cancer tissue subtyping. The proposed approach not only improves automated cancer classification, but it also enhances clinical trust.

Download

Share

COinS