Predictable Uncertainty-Aware Unsupervised Deep Anomaly Segmentation

Kazuki Sato, Kenta Hama, Takashi Matsubara, Kuniaki Uehara

Abstract: Image-based anomaly segmentation is a fundamental topic for image analysis. For medical use, it supports treatments via refined diagnosis and growth rate evaluation of tumors and lesions. Especially, an unsupervised training is expected to generalize to unknown anomalies. Probabilistic models have been used for this purpose, whereby these models are trained to maximize the likelihood of known samples and detect anomalous samples by assigning low likelihoods. Recent studies have proposed a probabilistic model based on deep neural networks (DNNs) called AEs and they achieved significant performance thanks to their flexibility. However, AEs are sensitive to complex structure (e.g., ridges and grooves of a brain) rather than semantic anomalies (e.g., tumors and lesions). We decomposed the approximated log-likelihood into two terms; predictable uncertainty and normalized error. We found that the former represents the complexity of structure. Hence, we propose the normalized error as a novel uncertainty-sensitive score by removing the predictable uncertainty. We evaluated our score by experiments with head magnetic resonance imaging (MRI) datasets and demonstrate the robustness of the proposed normalized error to data complexity.