Characterising borderline areas in bladder tumour grading with Bayesian graph neural networks

• Published in: 2024 IEEE International Symposium on Biomedical Imaging (ISBI) pp. 1 - 5
• Main Authors: Gao, Shangqi, Browning, Lisa, Alham, Nasullah Khalid, Protheroe, Andrew, Edwards, Keir, Hamblin, Josh, Rittscher, Jens, Verrill, Clare
• Format: Conference Proceeding
• Language: English
• Published: IEEE 27.05.2024
• Subjects: Bayesian graph neural networks; Bladder; Computational modeling; computational pathology; Decision making; Deep learning; Graph neural networks; Predictive models; tumour grading; Uncertainty; uncertainty quantification
• ISSN: 1945-8452
• DOI: 10.1109/ISBI56570.2024.10635485

Urothelial carcinoma is the most common bladder cancer whose grading is critical to clinical decision-making. The WHO 2004 grading system classifies urothelial carcinoma into either low grade or high grade, but sometimes cases sit on the border between grades. This makes assessment by the pathologist challenging but could potentially lead to under-treatment or overtreatment. The aim of this study was to use deep learning methods to identify and characterise borderline areas in whole slide images (WSIs) from bladder tumour cases. We constructed graphs on WSIs to accelerate computation, where positive unlabeled learning was utilized, accommodating the partial annotation strategy deployed in clinics. We used Bayesian deep learning for carcinoma classification, where we modeled the borderline as prediction uncertainty quantified by Bayesian graph neural networks. Our experiments showed promising performance of our approach in carcinoma detection and classification, with a potential use case to highlight and better characterise areas on the border for high grade and low grade to pathologists.