Explainable Artificial Intelligence in Non-Contrast Brain Computed Tomography Scan for Intracerebral Hemorrhage: A Scoping Review.

Abstract

Introduction: Artificial intelligence (AI) models applied to non-contrast brain computed tomography (CT) scan have demonstrated promising performance in hematoma detection, segmentation, and outcome prediction. Explainable artificial intelligence (XAI) has been proposed as a strategy to improve transparency and facilitate clinical integration. This study aimed to map and summarize XAI methods used in CT scan-based studies of intracranial hemorrhage (ICH), and to assess their validation approaches, transparency, and clinical relevance.

Method:  The scoping review was conducted in accordance with the PRISMA-ScR guidelines and was registered on the Open Science Framework (osf.io/5kxbt). In this review, original studies of adult and pediatric patients with spontaneous ICH that used AI or deep learning on non-contrast brain CT scan and included an explainability or interpretability method were eligible. Reviews, editorials, conference abstracts without full text, and studies without XAI components were excluded. PubMed, Embase, and Scopus were searched from inception to September 2025. Data were extracted on study design, imaging inputs, AI task, explainability technique, validation strategy, and reported clinical relevance.

Results: A total of twelve studies met inclusion criteria. Most investigations focused on hematoma detection, segmentation, or prediction of functional outcome or mortality. Gradient-based visualization techniques, particularly Grad-CAM and saliency maps, were the most commonly used explainability approaches. XAI outputs predominantly highlighted hematoma regions and perihematomal tissue; however, reporting and interpretation of explainability varied substantially across studies. External validation was infrequent, and few studies formally assessed alignment between XAI outputs and established radiological or clinical reasoning. The evidence was small, heterogeneous, and largely retrospective, with inconsistent reporting of explainability quality and limited external validation.

Conclusion: Explainable AI applied to non-contrast brain CT scan in ICH cases, is an evolving field with growing methodological diversity but limited standardization. While XAI techniques offer potential to enhance transparency and clinician confidence, current evidence remains insufficient to support routine clinical implementation. Future studies should emphasize standardized reporting, external validation, and clinically grounded evaluation of explainability outputs.