TLF: Triple learning framework for intracranial aneurysms segmentation from unreliable labeled CTA scans

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Intracranial aneurysms (IAs) are a significant health concern, and their diagnosis via computed tomography angiography (CTA) is complex.
• Deep neural networks (DNNs) show promise for medical image segmentation but require extensive labeled data, a bottleneck for IA analysis.
• Current methods struggle with the time-consuming and challenging annotation of brain CTA scans for IA detection.

Purpose of the Study:

• To develop a robust segmentation model for intracranial aneurysms (IAs) from brain CTA scans.
• To address the challenge of limited high-quality labeled data for training deep neural networks (DNNs).
• To propose a novel triple learning framework (TLF) integrating pseudo-supervised, contrastive, and confident learning.

Main Methods:

• Implemented a triple learning framework (TLF) combining pseudo-supervised, contrastive, and confident learning strategies.
• Utilized an enhanced mean teacher model and voxel-selective strategy for pseudo-supervised learning on noisy labels.
• Incorporated contrastive learning in the feature space and multi-scale confident learning for label correction and improved segmentation.

Main Results:

• The proposed TLF successfully learned a robust IA segmentation model from unreliable labeled brain CTA data.
• Experimental results on a large dataset demonstrated superior segmentation accuracy compared to existing state-of-the-art methods.
• The method effectively overcomes the limitations of data scarcity and annotation challenges in IA diagnosis.

Conclusions:

• The triple learning framework (TLF) offers an effective solution for segmenting intracranial aneurysms (IAs) using challenging brain CTA data.
• This approach significantly advances the potential of deep learning in medical image analysis, particularly for conditions requiring precise segmentation.
• The developed model shows promise for improving the efficiency and accuracy of IA diagnosis, aiding clinical decision-making.