Integrating Uncertainty Estimation with Deep Neural Networks for Improved Surgical Decision-Making

Abstract

Surgical margin assessment is critical to ensuring complete tumour excision while preserving healthy tissue. However, traditional intraoperative evaluation methods are often slow, subjective, and limited in scope. Rapid Evaporative Ionization Mass Spectrometry (REIMS), as implemented in the Intelligent Knife (iKnife), offers realtime molecular analysis of tissue during surgery and has shown promise in improving surgical outcomes. However, machine learning models used for classifying iKnife data often lack the ability to convey uncertainty. This poses a risk in high-stakes clinical environments and can undermine surgeon trust in Artificial Intelligence (AI) tools. This thesis investigates the integration of uncertainty estimation into deep learning models for surgical margin detection in Basal Cell Carcinoma (BCC). Three state-of-the-art uncertainty quantification techniques are used: Bayesian neural networks, deep ensembles, and evidential learning. These are implemented and evaluated using specialized architectures on both one-dimensional mass spectrometry signal data and two-dimensional image transformations. The models are assessed for predictive accuracy and uncertainty calibration quality, using metrics such as Expected Calibration Error and abstention-based trade-off analysis. Our findings demonstrate that uncertainty-aware models not only maintain high diagnostic accuracy but also effectively identify low-confidence predictions, enabling safer and more informed surgical decision-making. This work presents a comprehensive comparison of uncertainty methods on REIMS BCC data, and offers a foundation for deploying AI in the operating room with greater transparency and trust.