Uncertainty-Aware Deep Learning With Cross-Task Supervision for PHE Segmentation on CT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Perihematomal edema (PHE) volume is a key biomarker for spontaneous intracerebral hemorrhage (SICH)-associated diseases.
• Accurate PHE segmentation is crucial but challenging due to irregular shapes and low contrast on CT scans.
• Manual annotation is time-consuming and labor-intensive, hindering supervised deep learning applications.

Purpose of the Study:

• To develop an annotation-efficient deep learning framework for accurate PHE segmentation.
• To overcome the limitations of pixel-wise annotation requirements in current supervised methods.
• To leverage more accessible clinical data, such as slice-level PHE labels and pixel-wise SICH annotations.

Main Methods:

• Proposed a novel cross-task supervised framework for PHE segmentation.
• Utilized slice-level PHE labels to train a classifier generating pseudo PHE annotations (Class Activation Maps).
• Employed an uncertainty-aware corrective training strategy and a distance-aware loss for refining pseudo annotations and improving segmentation accuracy.

Main Results:

• The proposed framework achieved performance comparable to fully supervised methods for PHE segmentation.
• Demonstrated significant improvement over baseline methods trained solely with pseudo PHE labels.
• Validated the effectiveness of cross-task supervision in annotation-efficient deep learning for medical imaging.

Conclusions:

• The developed framework offers an effective solution for annotation-efficient PHE segmentation.
• Cross-task supervision presents a promising approach for similar challenges in medical image analysis.
• This methodology can potentially be adapted for other medical imaging segmentation tasks.