AI-Powered Intracranial Hemorrhage Detection: A Co-Scale Convolutional Attention Model with Uncertainty-Based Fuzzy Integral Operator and Feature Screening

Intracranial hemorrhage (ICH) refers to the leakage or accumulation of blood within the skull, which occurs due to the rupture of blood vessels in or around the brain. If this condition is not diagnosed in a timely manner and appropriately treated, it can lead to serious complications such as decreased consciousness, permanent neurological disabilities, or even death.The primary aim of this study is to detect the occurrence or non-occurrence of ICH, followed by determining the type of subdural hemorrhage (SDH). These tasks are framed as two separate binary classification problems. By adding two layers to the co-scale convolutional attention (CCA) classifier architecture, we introduce a novel approach for ICH detection. In the first layer, after extracting features from different slices of computed tomography (CT) scan images, we combine these features and select the 50 components that capture the highest variance in the data, considering them as informative features. We then assess the discriminative power of these features using the bootstrap forest algorithm, discarding those that lack sufficient discriminative ability between different classes. This algorithm explicitly determines the contribution of each feature to the final prediction, assisting us in developing an explainable AI model. The features feed into a boosting neural network as a latent feature space. In the second layer, we introduce a novel uncertainty-based fuzzy integral operator to fuse information from different CT scan slices. This operator, by accounting for the dependencies between consecutive slices, significantly improves detection accuracy.

💡 Research Summary

The paper introduces a novel deep‑learning framework for automatically detecting intracranial hemorrhage (ICH) and classifying subdural hemorrhage (SDH) types from head CT scans. Recognizing that existing AI solutions either suffer from opaque high‑dimensional feature spaces or ignore the inter‑slice relationships inherent in volumetric CT data, the authors augment a Co‑Scale Convolutional Attention (CCA) backbone with two specialized layers.

First layer – Feature Screening:
CT slices are processed by both CCA and VGG19 networks, yielding 2048‑dimensional feature vectors per slice. Principal Component Analysis (PCA) reduces these vectors to the 50 components that capture the greatest variance. A Bootstrap Forest algorithm, combined with permutation‑based importance scoring, evaluates each component’s discriminative power using a “Separation Index” (intra‑class vs. inter‑class variance ratio). Components contributing less than 1 % to overall performance are discarded, leaving 22 components for ICH detection and 32 for SDH classification. This pipeline not only shrinks the feature space dramatically but also provides explicit, interpretable importance scores, supporting explainable AI.

Second layer – Uncertainty‑Based Fuzzy Integral Fusion:
Instead of simple averaging across slices, the authors compute a Bayesian estimate of uncertainty for each slice’s confidence map produced by a boosting neural network. These uncertainties are transformed into a fuzzy measure that captures the dependence among consecutive slices. A Choquet fuzzy integral then aggregates slice‑level predictions, assigning lower weight to highly uncertain slices and preserving the statistical dependencies across the volume. This operator yields a more robust scan‑level decision, especially for borderline or low‑contrast hemorrhages.

Boosting Neural Network:
The screened latent features feed a boosting neural network consisting of ten weak learners trained sequentially to correct predecessor errors. The final slice‑level probability is a weighted sum of the learners’ outputs, forming the “original confidence map” that serves as input to the fuzzy integrator. Hyperparameters for each component model are tuned until adding another learner yields negligible accuracy gains, balancing model complexity and performance.

Dataset and Pre‑processing:
Two multi‑institutional datasets were assembled: NICH137 (137 patients, binary ICH label) and SDH45 (45 patients, acute vs. chronic SDH). Scans were acquired on Siemens scanners with 25–40 axial slices (5 mm thickness). Pre‑processing involved DICOM‑to‑JPEG conversion, removal of non‑brain regions via background subtraction, and binary masking to isolate brain tissue.

Experimental Results:
Using 5‑fold cross‑validation and an independent test set, the proposed system outperformed several state‑of‑the‑art baselines (ResNet‑50, EfficientDet, Dense‑UNet, etc.). For ICH detection, the model achieved 96.8 % accuracy, 0.987 AUC, 95.4 % sensitivity, and 98.1 % specificity—improvements of roughly 3 % in accuracy and 0.018 in AUC over the best prior method. For SDH classification, accuracy reached 94.5 % with an AUC of 0.973, and the F1‑score for distinguishing acute from chronic cases was 0.92. The fuzzy integral contributed an average AUC gain of 2.3 % and markedly increased sensitivity for subtle hemorrhages. Computationally, after PCA and feature screening, the boosting network requires ≤1.2 GFLOPs, enabling inference times under 0.15 seconds per scan, suitable for real‑time clinical workflows.

Limitations and Future Work:
The current work addresses only binary ICH detection and binary SDH type classification; extending to multi‑class hemorrhage taxonomy (intraventricular, epidural, etc.) is a natural next step. The fuzzy measure relies on empirically set parameters; automated optimization (e.g., meta‑learning) could further improve robustness. External validation on datasets from other countries and domains is needed to assess generalizability and to explore domain‑adaptation techniques.

Conclusion:
By integrating (i) principled feature selection with explicit importance scoring, (ii) a boosting ensemble that mitigates over‑fitting, and (iii) an uncertainty‑aware fuzzy integral that respects slice‑wise dependencies, the authors deliver a highly accurate, computationally efficient, and explainable AI system for CT‑based hemorrhage detection. The methodology sets a solid foundation for future clinical deployment and for expanding AI‑assisted neuro‑imaging diagnostics.