Multi-scale convolutional adversarial autoencoder-based anomaly detector for lesion edge localization on diffusion MRI

Motivation: Diffusion magnetic resonance imaging (dMRI) is an advanced technique that tracks water molecule movement in biological tissues. The spherical mean signal is invariant to fiber orientation distribution and depends solely on tissue microstructure, effectively characterizing the heterogeneity of human brain tissue. Goal(s): This study utilizes dMRI features to identify tumors by detecting abnormal water molecule signals, addressing the challenge of irregular tumor margins. Approach: We propose a Multi-Scale Convolutional Adversarial Autoencoder (MCAAE) model for effective localization of abnormal brain tissue margins. Results: Experiments indicate that this method successfully distinguishes between healthy and abnormal tissues, improving the delineation of irregular tumor margins. Impact: In this paper, we utilized dMRI features to identify tumors by analyzing the abnormal signals of microscopic water molecules. To address the challenge of irregular tumor margins, we proposed a method for detecting the abnormal probability of brain tissue.