Authors:
B. Manjunatha, K. Aravinda, M. Arunadevi Thirumalraj, S. Gopikha, M. Rehena Sulthana

Addresses:
Department of Mechanical Engineering, New Horizon College of Engineering, Bengaluru, Karnataka, India. Department of Electronics and Communications Engineering, New Horizon College of Engineering, Bengaluru, Karnataka, India. Department of Computer Science and Engineering, Karunya Institute of Technology and Science, Coimbatore, Tamil Nadu, India. Department of Computer Science and Business Management, Saranathan College of Engineering, Tiruchirappalli, Tamil Nadu, India. Department of Information Technology, St. Joseph's College of Engineering, Chennai, Tamil Nadu, India. School of Information Technology and Engineering, Melbourne Institute of Technology, Melbourne, Victoria, Australia.

Abstract:

Helpful tools for pathologists include the ability to identify and segment histological areas of diagnostic interest automatically. The challenge with segmentation methods is that obtaining pixel-level annotations for Whole-Slide images (WSI) is both time-consuming and costly. As a solution, weakly supervised approaches have been developed to leverage image-level annotations. But none of these methods have been modified to handle WSIs as far as researchers are aware. This research proposes a region-aware multi-task learning framework (RMTL-Net) to detect benign or malignant prostate tumours in images and simultaneously segment their regions. This model includes a regional attention (RA) module that uses the predicted likelihood maps to automatically train the classifier to learn category-sensitive information in the tumour, peritumoral, and background regions. By seamlessly fusing these regions, it improves feature representation and, in turn, the model's segmentation and classification performance. The Group Teaching Optimisation Algorithm (GTOA) model optimally selects the hyperparameter tuning. Using three publicly available prostate datasets, the research compared the proposed RA module with three state-of-the-art methods implemented recently and conducted extensive ablation experiments. 

Keywords: Tumor Classification; Whole-Slide Images; Regional Attention; Prostate Cancer; Regional-Attentive and Multi-Task Learning Framework; Group Teaching Optimization Algorithm.

Received on: 17/01/2025, Revised on: 24/03/2025, Accepted on: 27/05/2025, Published on: 06/12/2025

DOI: 10.69888/FTSHSL.2025.000511

FMDB Transactions on Sustainable Health Science Letters, 2025 Vol. 3 No. 4, Pages: 218-230

• Views : 212
• Downloads : 22

Download PDF