Incorporating Expert Prior Knowledge for Oral Lesion Recognition

Camille Besombes, Adeetya Patel, Sreenath Arekunnath Madathil

Keywords: Uncertainty Quantification, Model Calibration, Expert Knowledge, Oral Lesion Recognition

TL;DR: Our paper proposes a method to incorporate expert knowledge during training for medical image recognition.

Abstract: External information may improve predictive accuracy and uncertainty in medical image recognition. For example in oral lesion recognition, some lesion types are implausible to occur at certain anatomical locations. We propose a strategy to induce the prior knowledge about such correlations using an additional loss term that optimizes for plausible lesion types given an anatomical location. Our results suggests an improvement in model calibration, reduction in predicted number of implausible classes and improved uncertainty estimation for implausible predictions.