A Critical Review of Explainable Deep Learning for MRI-Based Brain Stroke Detection: From Accuracy to Trust

Abstract

Brain stroke is one of the most common neurological diseases characterized by significant mortality and long-term disabilities, so accurate and timely diagnosing is necessary for effective treatment. Deep learning (DL) models have shown promise in automating brain stroke detection and lesion segmentation tasks from magnetic resonance imaging (MRI). However, current literature lacks a discussion on how understandable and explainable the decisions made by AI are. This paper provides a critical review of the current state of explainable deep learning (DL) models for automated brain stroke detection via MRI analysis. In this review, we consider fourteen scientific papers dedicated to the topic of XAI in MRI-based brain stroke lesion detection in the period from 2020 to 2025. We analyze those articles based on the following criteria: presence of XAI methods, quantification of explanations validity, public access to the datasets, metrics used for evaluation, and clinical validation. We highlight major gaps in XAI research in the field which include lack of XAI methods use, absence of quantitative comparison of explanation maps to lesion segmentation performed by radiologists, private datasets, and insufficient clinical validation.