Challenging Current Semi-Supervised Anomaly Segmentation Methods for Brain MRI

TL;DR

This paper demonstrates that simple thresholding of FLAIR MRI scans outperforms several machine learning models in semi-supervised brain anomaly segmentation, challenging current ML-based approaches.

Contribution

The study shows that thresholding FLAIR scans can surpass ML models in brain anomaly segmentation, questioning the reliance on complex algorithms.

Findings

Thresholding FLAIR scans yields higher Dice scores.

Thresholding outperforms ML models in precision-recall metrics.

Simple thresholding is more effective than current ML approaches.

Abstract

In this work, we tackle the problem of Semi-Supervised Anomaly Segmentation (SAS) in Magnetic Resonance Images (MRI) of the brain, which is the task of automatically identifying pathologies in brain images. Our work challenges the effectiveness of current Machine Learning (ML) approaches in this application domain by showing that thresholding Fluid-attenuated inversion recovery (FLAIR) MR scans provides better anomaly segmentation maps than several different ML-based anomaly detection models. Specifically, our method achieves better Dice similarity coefficients and Precision-Recall curves than the competitors on various popular evaluation data sets for the segmentation of tumors and multiple sclerosis lesions.