Supervised and semi-supervised 3D organ localisation in CT images combining reinforcement learning with imitation...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Computer-aided diagnostics heavily relies on analyzing specific regions of interest (ROIs) in medical scans.
• Deep learning models excel in this but require extensive annotated data, which is often scarce in medical imaging.
• Existing methods for organ localization face challenges due to data limitations.

Purpose of the Study:

• To develop and evaluate an approach for localizing and detecting multiple organs using supervised and semi-supervised learning (SSL).
• To address the challenge of limited annotated data in medical image analysis.
• To improve the efficiency of computer-aided diagnostic systems.

Main Methods:

• The study presents a novel method for organ localization and detection based on supervised and semi-supervised learning.
• The approach builds upon prior work in localizing the thoracic and lumbar spine in CT images.
• Six bounding boxes for organs of interest are generated and fused into a single bounding box.

Main Results:

• Experiments demonstrated successful localization of the spleen, left kidney, and right kidney in CT images.
• The semi-supervised learning (SSL) approach effectively addressed data limitations, requiring significantly less data and fewer annotations compared to state-of-the-art methods.
• SSL performance was evaluated using various labeled/unlabeled data ratios (30:70, 35:65, 40:60).

Conclusions:

• Semi-supervised learning (SSL) offers a viable and effective alternative for organ localization in medical imaging, particularly where annotated data is difficult to obtain.
• The proposed method shows promise for improving computer-aided diagnostics by overcoming data scarcity challenges.
• The findings highlight the potential of SSL in enhancing the development of robust medical imaging analysis tools.