Uncertainty Modelling for Tumour Cellularity Estimation in Histopathology Using Deep Learning

Abstract

Tumour cellularity (TC) is an important metric used in the cancer treatment journey, from monitoring the therapeutic response to guiding subsequent treatment decisions. It can be defined as the area occupied by malignant cells in a Region of Interest. Although deep learning (DL)-based automated systems have gained considerable momentum in medical image analysis, there are still challenges associated with efficient TC computation. The scope of performance improvement by comprehensively addressing issues of aleatoric uncertainty in TC labels and epistemic uncertainty in model parameters due to limited data remains unexplored. In this research, a novel regression-based framework is developed to address these two types of uncertainties for efficient computation of TC. While a loss function is introduced to handle aleatoric uncertainty, the epistemic uncertainty is handled by appropriately using Monte Carlo dropout and training the DL model in two stages by skilfully utilizing limited data. First, a novel task-specific pre-training is done in a semi-supervised framework to capture fine-grained cell-level features. Then, it is fine-tuned for effective TC computation. The experimental results show effective performance with different encoders, demonstrating robustness. Quantitative and qualitative results demonstrate the potential of the framework to serve as a robust tool for clinical decision support in cancer. Achieving a PK score of 0.95 and an R2 score of 0.88, it outperforms state-of-the-art histopathology-specific foundation models and other related methods w.r.t. both metrics and parameters. Thus, our framework offers a reliable and scalable solution for improved automated tumour cellularity estimation and for reducing the workload of pathologists.