# Instance-level quantitative saliency in multiple sclerosis lesion segmentation

## TL;DR

This paper introduces new explainable AI methods to understand how models detect and outline brain lesions in multiple sclerosis patients using MRI scans.

## Contribution

The paper introduces two novel XAI methods for generating quantitative instance-level explanations in semantic segmentation.

## Key findings

- The proposed saliency maps showed models rely more on FLAIR than MPRAGE for white matter lesion segmentation.
- Peak saliency values differ significantly between true and false predictions, suggesting potential for error identification.
- The methods can be applied to any architecture to improve performance and justify decisions for specific lesions.

## Abstract

In recent years, explainable methods for artificial intelligence (XAI) have tried to reveal and describe models’ decision mechanisms in the case of classification and even for segmentation. However, XAI methods for semantic segmentation and in particular for single specific instances (e.g. one given lesion among others of the same class in medical imaging) have yet to be developed to understand what drove the detection and contouring of the latter, which is crucial for all multi-lesional diseases. We proposed instance-level explanation maps for semantic segmentation extending both SmoothGrad and Grad-CAM++ methods and yielding quantitative instance saliency for the former. The instance-level methods were applied to the segmentation of white matter lesions (WML), a magnetic resonance imaging (MRI) biomarker in multiple sclerosis (MS). 687 patients diagnosed with MS for a total of 4023 FLAIR and MPRAGE MRI scans were collected at the University Hospital of Basel, Switzerland. WM lesion masks were annotated by four expert clinicians on baseline and follow-up imaging. Three deep learning networks—a 3D U-Net, nnU-Net, and Swin UNETR—were trained and tested on these data (test normalized Dice score, respectively of 0.71, 0.78, 0.80; true positive rate of 79%, 78%, and 85%; false discovery rate of 37%, 38%, and 36%; false negative rate of 20%, 22%, and 14%), then saliency maps were computed. Consistent with clinical practice, the proposed instance saliency maps revealed that the model relied more on FLAIR than MPRAGE to segment WMLs, with positive saliency values inside a lesion and negative in its neighborhood. FLAIR hyperintensity combined with healthy WM around the lesion border was required for their detection. Beyond the aforementioned sanity checks, we observed that peak values of the generated saliency maps presented distributions that differ significantly between TP, FN, FP and TN predictions, suggesting that the quantitative nature of the proposed saliency could be used to identify errors. In conclusion, we introduced two XAI methods to generate quantitative instance-level explanations in semantic segmentation. The proposed XAI maps can be applied to any architecture and could serve as a basis to (i) improve model performance (e.g. reducing FPs), (ii) optimize their internal architecture (e.g. patch size), and (iii) justify the model’s decisions to the end users, which are contextualized to a specific lesion instance of interest.