Vision Transformer for Intracranial Hemorrhage Classification in CT Scans Using an Entropy-Aware Fuzzy Integral Strategy for Adaptive Scan-Level Decision Fusion

TL;DR

This work addresses the critical task of multi-class intracranial hemorrhage (ICH) subtype classification from brain CT by introducing a Pyramid Vision Transformer (PVT) framework augmented with SHAP-driven feature selection and an entropy-aware fuzzy integral for scan-level fusion. Key contributions include a comprehensive, expert-labeled dataset from two Tehran centers with careful patient-level splits, a preprocessing pipeline to focus on brain tissue, and a feature space pruned by SHAP-informed selection and PCA. The method combines a boosting neural network for slice-level predictions with an entropy-based Choquet integral fusion across slices, yielding robust, accurate scan-level diagnoses that outperform several state-of-the-art architectures. The proposed approach demonstrates improved reliability and computational efficiency, supporting scalable AI-assisted ICH classification in clinical settings, and lays groundwork for future multi-modal and explainable deployments.

Abstract

Intracranial hemorrhage (ICH) is a critical medical emergency caused by the rupture of cerebral blood vessels, leading to internal bleeding within the skull. Accurate and timely classification of hemorrhage subtypes is essential for effective clinical decision-making. To address this challenge, we propose an advanced pyramid vision transformer (PVT)-based model, leveraging its hierarchical attention mechanisms to capture both local and global spatial dependencies in brain CT scans. Instead of processing all extracted features indiscriminately, A SHAP-based feature selection method is employed to identify the most discriminative components, which are then used as a latent feature space to train a boosting neural network, reducing computational complexity. We introduce an entropy-aware aggregation strategy along with a fuzzy integral operator to fuse information across multiple CT slices, ensuring a more comprehensive and reliable scan-level diagnosis by accounting for inter-slice dependencies. Experimental results show that our PVT-based framework significantly outperforms state-of-the-art deep learning architectures in terms of classification accuracy, precision, and robustness. By combining SHAP-driven feature selection, transformer-based modeling, and an entropy-aware fuzzy integral operator for decision fusion, our method offers a scalable and computationally efficient AI-driven solution for automated ICH subtype classification.

Figures (8)

• Figure 1: Illustrative examples of different types of brain hemorrhages in CT scans. The hemorrhagic regions are highlighted in red. (a) IPH, (b) IVH, (c) EDH, (d) SAH, (e) SDH.
• Figure 2: Examples of CT scan slices that were removed during preprocessing. These slices were eliminated because of lack of relevant brain tissue, or poor image quality, ensuring that only the most informative slices contribute to the classification process.
• Figure 3: An overview of the proposed scan-level classification framework for ICH detection.
• Figure 4: SHAP-based feature importance ranking. The ranked components correspond to the PVT model, highlighting the most influential features contributing to the classification of ICH subtypes.
• Figure 5: Comparison of computational time across different processing methods in the PVT-based model.