Explainable human‐in‐the‐loop healthcare image information quality assessment and selection

Abstract

AbstractSmart healthcare applications cannot be separated from healthcare data analysis and the interactive interpretability between data and model. A human‐in‐the‐loop active learning approach is introduced to reduce the cost of healthcare data labelling by evaluating the information quality of unlabelled medical data and then screening the high informative samples. Specifically, focused on the extracted feature tensor of medical images, a new evaluation metric was proposed, called local feature information distance. Then, the whole feature information distance between two images can be obtained by seeking optimal matching of local feature information correlation based on the feature spatial traversal. Further, the feature information distance entropy (FIDE) method was proposed from the perspective of feature spatial distribution, which outperformed other related works. Many comparison and ablation experiments were carried out; the results showed that the proposed method can well distinguish the information quality of the unlabelled samples even facing different datasets. The stability, robustness, and generalisation of the proposed method were verified. Thus, this study provides some inspirations for healthcare data analysis and efficient management, such as eliminating redundancy and selective annotation.