Utilizing shallow features and spatial context for weakly supervised intracerebral hemorrhage segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Weakly-supervised semantic segmentation (WSSS) for intracerebral hemorrhage (ICH) diagnosis is limited by the lack of precise location information from image-level labels.
• Developing effective WSSS methods is crucial for improving automated ICH segmentation and diagnosis.

Purpose of the Study:

• To develop a novel method for enhancing intracerebral hemorrhage segmentation using weak image-level labels.
• To improve the accuracy of target localization and contour detection in ICH segmentation.

Main Methods:

• Proposed a Shallow-Feature Class Activation Map (CAM) module to leverage fine-grained shallow features for accurate localization.
• Introduced a Spatial Context Aware (SCA) module to incorporate spatial context and complete hemorrhage segmentation.
• Validated the method on the Brain Hemorrhage Segmentation Dataset (BHSD) and the CT Images for Intracranial Hemorrhage Detection and Segmentation Dataset (BCIHM).

Main Results:

• The proposed method significantly improved ICH segmentation accuracy, increasing mean Intersection over Union (mIoU) from 52.5% to 69.8% (BHSD) and 50.1% to 68.9% (BCIHM).
• Achieved superior performance compared to other WSSS methods, reaching 88% and 86% of fully supervised U-Net performance.
• Demonstrated significant reductions in missed and false positive localizations.

Conclusions:

• The novel WSSS method effectively improves intracerebral hemorrhage segmentation by accurately matching lesion location and contours.
• The method reduces missed and false positive localizations, thereby decreasing the workload for radiologists in creating pixel-level datasets.