A proposal of Texture Features for interactive CTA Segmentation by Active Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of the abdominal aorta is crucial for diagnosing and managing Abdominal Aortic Aneurysms (AAA).
• Manual segmentation is time-consuming and operator-dependent.
• Developing automated or semi-automated systems can improve efficiency and consistency.

Purpose of the Study:

• To develop an interactive image segmentation system for rapid volumetric segmentation of the abdominal aorta.
• To minimize human operator intervention in the segmentation process.
• To apply the system to the segmentation of thrombus in Computed Tomography Angiography (CTA) data of AAA patients.

Main Methods:

• Utilizing an Active Learning approach combined with enhanced image texture features (Gray Level Co-occurrence Matrix and Local Binary Patterns).
• Employing a Random Forest classifier trained on labeled voxels.
• Iterative retraining of the classifier based on human operator selection of uncertain voxels.
• Incorporating a priori knowledge of target structure shape to refine detections.

Main Results:

• Preliminary experiments show promising segmentation results on diverse CT datasets.
• The used texture features effectively describe local variations in AAA thrombus.
• The system demonstrates the capacity to provide useful information for automated classification.

Conclusions:

• The developed interactive system shows potential for efficient and accurate abdominal aorta and thrombus segmentation.
• Active learning with texture features offers a viable strategy for reducing manual effort in medical image analysis.
• Further validation on larger datasets is warranted to confirm clinical utility.