Contents I. Introduction
Interpretability, reliability and explainability play an essential role in Deep Learning (DL), particularly when DL tools are exploited in critical and sensitive scenarios such as clinical and research medical applications. These three principles simultaneously provide proper guidance and open new possibilities for DL. Despite the limitations imposed by it, developing reliable and explainable DL systems consequently produces a richness of information that can be exploited to gain new perspectives on the task at hand. Of the many medical-related fields in which DL is slowly making itself at home, neuroscience is certainly noteworthy, whether it is for image analysis, classification, feature extraction or data clustering. Specifically, a major effort is being advanced in understanding Autism Spectrum Disorder (ASD), with the Autism Brain Imaging Data Exchange (ABIDE) dataset being a key source of anatomical and functional data useful in shedding light on this important and increasingly prevalent issue [1]. Many have already gave their contribution on this pressing matter through the lens of DL such as Heinsfeld et al. [2], Yang et al. [3], Li et al. [4], Epalle et al. [5] and Prasad et al. [6] whether attempting to classify subjects based on various available features, interpret and study the dataset and/or use eXplainable AI (XAI) and classical methods to identify autism biomarkers. A comprehensive review of AI applications in ASD diagnosis is available from Valizadeh et al. [7]. Bayesian neural netowrks (BNNs) are a family of DL architectures which are collecting interesting results in the healthcare domain [8]. A BNN is at its core a probabilistic model augmented with a neural network as a universal function approximator [9]. This particular setup gives a BNN the capability of producing stochastic outputs (i.e. repeating the inference with the same model and the same input results in different outputs), thus rendering it possible to assess epistemic uncertainty (uncertainty of the model) in the form of prediction confidence. BNN and XAI can both be used to aid diagnosis and have potential as knowledge-discovery tools; however, to the best of our knowledge, today there are no works which combine them to support neuroscience research in ASD biomarkers detection. XAI offers different strategies to explain DL models such as BNNs. Local techniques generate a specific explanation for every predicted outcome, while global methods aim to provide a “big picture” of the entire model behaviour. Attribution-based methods are XAI approaches widely-used to generate an explanation for every prediction returned by a blackbox model, that is, a local explanation of the AI-model [10], [11]. They generally work by assigning a score to the input features according to a certain importance criteria in the prediction process. There are currently different methods to score the importance of the input features, these includes Layerwise Relevance Propagation (LRP), which reported both theoretical and empirical advantages in representing the computation performed by the model [12]. In this work, the probabilistic nature of BNNs will be exploited to obtain confidence estimates on the predictions and compute stochastic outputs from the explanations produced using the LRP technique. The aim is to integrate explainability and bayesian neural networks to develop a classification framework able to (i) provide model's uncertainty metrics together with predictions; (ii) explore possible biomarkers candidates in ASD exploiting the probabilistic nature of the outputs.
