Towards Reliable Colorectal Cancer Polyps Classification via Vision Based Tactile Sensing and Confidence-Calibrated Neural Networks
This work addresses reliability issues in AI-based colorectal cancer diagnosis, which is critical for medical applications, though it is incremental as it builds on existing methods.
The study tackled the problem of overconfident AI outputs in colorectal cancer polyp classification by proposing a confidence-calibrated residual neural network, achieving improved reliability through temperature scaling and testing with noisy/blurred inputs.
In this study, toward addressing the over-confident outputs of existing artificial intelligence-based colorectal cancer (CRC) polyp classification techniques, we propose a confidence-calibrated residual neural network. Utilizing a novel vision-based tactile sensing (VS-TS) system and unique CRC polyp phantoms, we demonstrate that traditional metrics such as accuracy and precision are not sufficient to encapsulate model performance for handling a sensitive CRC polyp diagnosis. To this end, we develop a residual neural network classifier and address its over-confident outputs for CRC polyps classification via the post-processing method of temperature scaling. To evaluate the proposed method, we introduce noise and blur to the obtained textural images of the VS-TS and test the model's reliability for non-ideal inputs through reliability diagrams and other statistical metrics.