Concurrent Ischemic Lesion Age Estimation and Segmentation of CT Brain Using a Transformer-Based Network.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Accurate timing of stroke onset is critical for effective patient management.
• Interpreting Computed Tomography (CT) scans to determine lesion age is challenging due to subtle visual cues.
• Existing automated methods treat lesion segmentation and age estimation independently, missing their synergistic potential.

Purpose of the Study:

• To develop and evaluate a novel deep learning network for concurrent segmentation and age estimation of cerebral ischemic lesions.
• To leverage the complementary relationship between lesion segmentation and age estimation for improved accuracy.
• To address the limitations of current automated approaches in stroke imaging analysis.

Main Methods:

• Proposed a novel end-to-end multi-task transformer-based network for concurrent segmentation and age estimation.
• Utilized gated positional self-attention and CT-specific data augmentation for enhanced feature learning.
• Incorporated uncertainty estimation using quantile loss to generate a probability density function for lesion age.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.933 for classifying lesion ages ≤ 4.5 hours, outperforming conventional methods (0.858).
• Demonstrated superior performance compared to task-specific state-of-the-art algorithms.
• Validated the model on a clinical dataset of 776 CT images from two medical centers.

Conclusions:

• The proposed multi-task learning approach effectively integrates lesion segmentation and age estimation for improved stroke diagnosis.
• The transformer-based network shows significant potential for enhancing automated analysis of acute ischemic lesions in brain CT scans.
• This method offers a promising advancement in time-sensitive stroke care by providing more accurate lesion age determination.