Supervisor

Suitable for

Abstract

Medical data acquired in various modalities (CT, MRI, photograph) and of various anatomical parts is used in clinical decision making. Increasingly, machine learning methods are used in classification or segmentation tasks. Yet neural networks are known to be miscalibrated and often provide overconfident uncertainty estimates. The goal of this project is to evaluate the impact of different pretraining strategies (e.g., contrastive learning, self-supervised learning) and different fine-tuning strategies (e.g., data augmentation, test-time augmentation, label smoothing) on model calibration.