Unified ICH quantification and prognosis prediction in NCCT images using a multi-task interpretable network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning advancements have spurred interest in Computer-Aided Diagnosis (CAD) for spontaneous IntraCerebral Hematoma (ICH) detection using Non-Contrast head Computed Tomography (NCCT).
• Existing methods face challenges including time-intensive manual evaluation, high costs for patient-level predictions, and the need for both high accuracy and interpretability.

Purpose of the Study:

• To propose a novel multi-task deep learning framework to address the limitations of current ICH detection methods.
• To enhance the efficiency, accuracy, and interpretability of ICH diagnosis in emergency medicine.

Main Methods:

• A multi-task framework with upstream and downstream components was developed.
• An upstream, weight-shared module was trained for robust feature extraction via multi-task regression and classification.
• A downstream component utilized two separate heads for distinct regression and classification tasks.

Main Results:

• The proposed multi-task framework demonstrated superior performance compared to single-task frameworks.
• The model's interpretability was validated using Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps.
• The framework offers a more efficient and accurate approach to ICH volume evaluation.

Conclusions:

• The multi-task deep learning framework effectively overcomes challenges in ICH detection and volume evaluation.
• The proposed method achieves better performance and interpretability than traditional single-task approaches.
• This framework holds significant potential for improving emergency medicine diagnostics for spontaneous ICH.