Influence based explainability of brain tumors segmentation in multimodal Magnetic Resonance Imaging

TL;DR

We address explainability in multiclass brain-tumor segmentation by extending the TracIn influence-based xAI method to region-wise, per-class explanations using a 2D UNet on multimodal MRI from BraTS19. The approach yields local proponents/opponents per tumor class and couples these signals to latent feature maps via a faithfulness metric, enabling both local and global explanations for radiologists. Results show class-specific influence patterns and a meaningful correlation between TracIn scores and the network's internal representations, supporting faithful interpretation and potential feature selection. This work advances clinically usable xAI for segmentation tasks by providing quantifiable, region-aware explanations with practical interpretability for clinicians.

Abstract

In recent years Artificial Intelligence has emerged as a fundamental tool in medical applications. Despite this rapid development, deep neural networks remain black boxes that are difficult to explain, and this represents a major limitation for their use in clinical practice. We focus on the segmentation of medical images task, where most explainability methods proposed so far provide a visual explanation in terms of an input saliency map. The aim of this work is to extend, implement and test instead an influence-based explainability algorithm, TracIn, proposed originally for classification tasks, in a challenging clinical problem, i.e., multiclass segmentation of tumor brains in multimodal Magnetic Resonance Imaging. We verify the faithfulness of the proposed algorithm linking the similarities of the latent representation of the network to the TracIn output. We further test the capacity of the algorithm to provide local and global explanations, and we suggest that it can be adopted as a tool to select the most relevant features used in the decision process. The method is generalizable for all semantic segmentation tasks where classes are mutually exclusive, which is the standard framework in these cases.

Figures (7)

• Figure 1: Example of a patient from the BraTS19 dataset. The upper row of images shows a slice of an MRI scan in the four acquisition modalities, the bottom row shows the segmented regions in the four classes
• Figure 2: Scheme of the proposed methodology for TracIn calculation in the multiclass segmentation task considered. TracIn is computed separately for each region of the test image $z'$ predicted to be of class $i$ ($z'_i$). Each train image $z$ is also split into regions according to the network prediction ($z'_j$). A threshold on the network output is applied to eliminate pixels with greater prediction errors. For a given $z'_i$ a list of proponents and opponents is provided
• Figure 3: Scheme of the feature importance calculation. For first we extract the filters from the last hidden layer of the UNet and we measure the feature maps only on the selected tissue of the images on the training and test set. We then take the maximum value on the obtained maps and we evaluate the similarity metric (S) as described in the text. Taking the average of S with respect to the test set we define the feature importance for filter $k$ and class $i$ ($FI(z_j)_{k,i}$)
• Figure 4: Train-influence matrix for the BraTS19 training set. The diagonal terms ($i$ = $j$) represent the influence that the region of $z$, predicted to be of class $i$, has on the predicted region of the same class of $z'$, while off-diagonal terms quantify the influence of regions of $z$ predicted to be of classes different from $i$
• Figure 5: Distribution of the class of the top 10 proponents for the three train-influence matrices