Multi-Scale Self-Supervised Consistency Training for Trustworthy Medical Imaging Classification.

Modern neural network models have demonstrated exceptional classification capabilities comparable to human performance in various medical diagnosis tasks. However, their practical application in real-world medical scenarios is hindered by an issue known as miscalibration, where these sophisticated tools inaccurately estimate their own prediction confidence, compromising their rustworthiness. To address this challenge, we propose a novel neural network calibration framework that utilizes multi-scale input images and integrates self-supervised consistency enforcement during training. Our experimental results demonstrate the significant enhancement of neural network calibration, concomitant with improvements in model classification performance. Furthermore, the proposed method exhibits the capacity to cultivate more robust feature spaces. Importantly, our approach is a general-purpose solution that is applicable to any imaging modalities. The proposed method can also be combined with other neural network calibration techniques to achieve further performance refinement. This research contributes a valuable tool for augmenting the reliability and trustworthiness of neural network models in diverse medical contexts.

Download paper here