Unsupervised lesion detection via image restoration with a normative prior

Xiaoran Chen, Suhang You, Kerem Can Tezcan, Ender Konukoglu

Abstract: Highlights•Lesion detection with MAP-based image restoration improves the detection accuracy to 0.83 (AUC) and 0.45 (Dice) on BRATS17, and to 0.79 (AUC) and 0.12 (Dice) on ATLAS datasets, outperforming existing methods.•A threshold that distinguishes the lesions and the healthy anatomy can be determined using the proposed method with only the training set.•Detection of large lesions with large intensity alternation is more accurate than that of small lesions that have similar features to healthy anatomical structures.