Simulated Quantum Mechanics-Based Joint Learning Network for Stroke Lesion Segmentation and TICI Grading.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate stroke lesion segmentation and thrombolysis in cerebral infarction (TICI) grading are crucial for diagnosis.
• Previous methods often focused on single tasks, neglecting the relationship between segmentation and grading.
• This limitation hinders comprehensive stroke assessment.

Purpose of the Study:

• To develop a joint learning network for simultaneous stroke lesion segmentation and TICI grade assessment.
• To address the correlation and heterogeneity between these two tasks.
• To improve the accuracy and efficiency of auxiliary stroke diagnosis.

Main Methods:

• Proposed a simulated quantum mechanics-based joint learning network (SQMLP-net) with a single-input double-output architecture.
• Employed a shared encoder for extracting spatial and semantic information for both tasks.
• Utilized a novel joint loss function to optimize intra- and inter-task weights.

Main Results:

• SQMLP-net achieved state-of-the-art performance on the ATLAS R2.0 dataset, with Dice score of 70.98% and accuracy of 86.78%.
• The proposed method outperformed single-task and existing advanced methods.
• Analysis revealed a negative correlation between TICI grading severity and stroke lesion segmentation accuracy.

Conclusions:

• The SQMLP-net effectively performs simultaneous stroke lesion segmentation and TICI grading.
• Joint learning enhances diagnostic accuracy by leveraging the relationship between the two tasks.
• This approach offers a promising advancement in auxiliary stroke diagnosis.