Dual-task vision transformer for rapid and accurate intracerebral hemorrhage CT image classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neurology

Background:

• Intracerebral hemorrhage (ICH) is a critical condition requiring immediate diagnosis and treatment.
• Brain CT imaging is standard for ICH diagnosis, but analysis can be challenging due to image complexity and radiologist shortages.
• Rapid, accurate ICH assessment is vital for effective patient management and improved outcomes.

Purpose of the Study:

• To develop and evaluate an automated system for classifying intracerebral hemorrhage (ICH) and its subtypes from CT images.
• To address the challenges of timely ICH diagnosis in clinical settings through advanced AI.
• To create a robust deep learning model capable of analyzing real-world ICH CT datasets.

Main Methods:

• A real-world dataset of CT images was curated for normal vs. ICH classification and ICH subtype classification (Deep, Subcortical, Lobar).
• A novel neural network architecture, the dual-task vision transformer (DTViT), was proposed, utilizing Vision Transformer (ViT) encoders for feature extraction.
• The DTViT model incorporated two multilayer perception (MLP)-based decoders for simultaneous ICH detection and hemorrhage location classification.

Main Results:

• The DTViT model demonstrated strong performance in classifying ICH presence and differentiating between Deep, Subcortical, and Lobar hemorrhage locations.
• Experimental results validated the effectiveness of the DTViT framework on the collected real-world test dataset.
• The attention mechanisms within the ViT encoder facilitated effective feature extraction from complex CT images.

Conclusions:

• The proposed dual-task vision transformer (DTViT) offers a promising automated solution for intracerebral hemorrhage (ICH) diagnosis using CT imaging.
• DTViT's ability to simultaneously classify ICH presence and location can expedite treatment planning in emergency settings.
• This AI-driven approach has the potential to alleviate the burden on specialist radiologists and improve diagnostic efficiency for ICH.