This research article questions the fairness and interpretability of deep learning classification results obtained on CT scans. The authors point out that most deep learning-based classification attempts solely focus on high accuracy scores, without considering interpretability or patient-wise separation of training and test data. This leads to misleading accuracy rates and irrelevant feature learning, reducing the real-life usability of these models. The authors argue that deep neural networks perform better when trained with images of patients that are strictly isolated from the validation and testing patient sets.
Publication date: 25 Sep 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2309.12632
