Interpretable deep learning as a means for decrypting disease signature in multiple sclerosis

Abstract

Objective. The mechanisms driving multiple sclerosis (MS) are still largely unknown, calling for new methods allowing to detect and characterize tissue degeneration since the early stages of the disease. Our aim is to decrypt the microstructural signatures of the Primary Progressive versus the Relapsing-Remitting state of disease based on diffusion and structural magnetic resonance imaging data. Approach. A selection of microstructural descriptors, based on the 3D-Simple Harmonics Oscillator Based Reconstruction and Estimation and the set of new algebraically independent Rotation Invariant spherical harmonics Features, was considered and used to feed convolutional neural networks (CNNs) models. Classical measures derived from diffusion tensor imaging, that are fractional anisotropy and mean diffusivity, were used as benchmark for diffusion MRI (dMRI). Finally, T1-weighted images were also considered for the sake of comparison with the state-of-the-art. A CNN model was fit to each feature map and layerwise relevance propagation (LRP) heatmaps were generated for each model, target class and subject in the test set. Average heatmaps were calculated across correctly classified patients and size-corrected metrics were derived on a set of regions of interest to assess the LRP contrast between the two classes. Main results. Our results demonstrated that dMRI features extracted in grey matter tissues can help in disambiguating primary progressive multiple sclerosis from relapsing-remitting multiple sclerosis patients and, moreover, that LRP heatmaps highlight areas of high relevance which relate well with what is known from literature for MS disease. Significance. Within a patient stratification task, LRP allows detecting the input voxels that mostly contribute to the classification of the patients in either of the two classes for each feature, potentially bringing to light hidden data properties which might reveal peculiar disease-state factors.

MiFL: Multi-Input Neural Networks in Federated Learning

“…One of the strengths of ML models is the capability to capture hidden relationships among complex multi-dimensional data. They have been explored for several tasks, inter-alia, disease classification [1][2][3] , human body segmentation [4][5][6][7] , the definition of diagnostic scores 8 , drug-discovery 9,10 and data augmentation through the generation of synthetic samples [11][12][13][14] . Despite most of these examples leveraging medical images as the preferred data format, a variety of data types are collected by hospitals, clinical laboratories, and other healthcare institutions.…”

Explainable AI (XAI) In Biomedical Signal and Image Processing: Promises and Challenges

“…Among the most exploited methods are those relying on heatmaps revealing which parts of the input played a relevant role in determining the output. An example of a comprehensive solution was proposed in [10], where LRP was used to interpret the results of a patient stratification task aiming at distinguishing two multiple sclerosis phenotypes (progressive versus relapsing-remissing) and validation included the assessment of the role of confounds on the classification outcomes as well as association studies of LPR maps with microstructural descriptors that were previously shown to be significantly different in the two phenotypes of disease. Another example is [11] where the objective was to contrast feature-based and example-based explainable AI techniques using the Heart Disease Dataset from UCI (https://www.kaggle.com/cherngs/ heart-disease-clevelanduci) to highlight the respective pros and cons.…”

Deciphering multiple sclerosis disability with deep learning attention maps on clinical MRI

“…However, the use of DL techniques to predict disease progression is still largely unexplored. Some recent studies have presented promising results for future prediction of disability ( Roca et al, 2020 , Storelli et al, 2022 , Tousignant et al, 2019 ) and cross-sectional patient stratification ( Cruciani et al, 2021 , Marzullo et al, 2019 ).…”

“…New tendencies in the field have also put emphasis on producing explainable DL-based techniques, with the goal of disentangling the reason behind DL decisions, therefore providing additional information that could be extremely useful to the end users enhancing data insights ( Bach et al, 2015 , Simonyan et al, 2013 , Springenberg et al, 2014 ). This task of deciphering the “black box” of DL-based models in MS has recently been studied for disease diagnosis ( Eitel et al, 2019 , Lopatina et al, 2020 ) and MS phenotype signature decrypting ( Cruciani et al, 2021 ). All these studies concluded that the use of the layer-wise relevance propagation (LRP)( Bach et al, 2015 ), among other (similar) methods, is the most promising tool for these analyses providing individual heatmaps for each subject, called attention maps, reflecting the voxel-specific relevance to the classification output, according to the DL model, in an easy and intuitive way.…”