AUTHORS: Spagnolo F., Gordaliza P.M., Lu P.J., Ocampo Pineda M., Chen X., Weigel M., Wynen M., Absinta M., Maggi P., Bach Cuadra M., Andrearczyk V., Depeursinge A., Granziera C. 4o

European Committee For Treatment And Research In Multiple Sclerosis, : , Copenhagen, September 2024

ABSTRACT

Introduction: Paramagnetic Rim Lesions (PRLs) are dark rim-like areas of focal damage, which are visible in the brains of Multiple Sclerosis (MS) patients using susceptibility-based magnetic resonance imaging. Recent studies have used Deep Learning (DL) for their classification, but these methods are deployed as black boxes. Explainable methods (XAI) can help build trust in these models and facilitate their adoption in clinical practice.

Objective/Aims: To better understand what drives the decision of a DL network employed to classify PRLs.

Methods: Two trained experts annotated 384 PRLs using FLuid Attenuated Inversion Recovery (FLAIR) and T2*-w Unwrapped Phase (UP) images from 124 MS patients (77 females, age 45.0 ± 13.1, median EDSS 2.0 [1.5-4.5]) collected at the Lausanne University Hospital (Lausanne, Switzerland) and at the Universitätsspital Basel (Basel, Switzerland). Patches of 28x28x28 from both images (176x240x256 FLAIR, 256x336x384 UP) were extracted around MS lesions (which were automatically segmented and manually corrected), intensity normalised (0-1) and fed as input to train a convolutional neural network (RimNet). RimNet, trained on FLAIR only, has shown a good sensitivity but a poorer specificity compared to the use of UP. 12 patients (78 PRLs) were used for inference. We generated an attention map for each input sequence with the XAI method Integrated gradients. First, the distributions of extreme values (maximum and minimum) in FLAIR- and UP-based XAI maps were compared. Then, we reported how frequently the extreme values were found in the manually segmented rims, which would mean that the model’s decisions are particularly influenced by voxel values in these regions. The Wilcoxon signed-rank was adopted to compare FLAIR- and UP-based XAI maps.

Results: RimNet reached a precision/recall of 0.80/0.86, classifying 67 True Positive (TP), 16 False Positive (FP), 11 False Negative (FN), and 988 True Negative (TN) examples. UP-based maps presented negative values within the rim, and positive in its close neighbourhood. Values’ distribution in FLAIR-based maps did not show a clear correspondence to the rim. A higher mean of extreme values was found in UP-based maps ([-0.297, 0.300] for TPs, while [-0.232, 0.189] in FLAIR), with a p-value < 0.05 except in TNs and FPs. In the second test on TPs and FNs, minimum values of UP-based maps were found most frequently on the rim (respectively >50% and >72% of the times), with peak values of -0.286 ± 0.127 in TPs and of -0.297 ± 0.127 in FNs.

Conclusion: XAI RimNet confirms that UP’s voxels are more important than FLAIR’s to classify PRLs. Predictions can be attributed to dark voxels within the rim in UP and, perhaps, to darker demyelinated lesion cores in FLAIR.

BibTex

Module: