SCL-Net: Structured Collaborative Learning for PET/CT Based Tumor Segmentation.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Current collaborative learning for medical image segmentation, often UNet-based, faces limitations in ensuring auxiliary classifier optimization improves target classifier performance.
• Independent supervision of auxiliary classifiers hinders effective knowledge transfer and collective improvement in segmentation tasks.

Purpose of the Study:

• To introduce a novel Structured Collaborative Learning (SCL) method to enhance medical image segmentation accuracy.
• To address the limitations of independent classifier supervision in collaborative learning frameworks.

Main Methods:

• Proposed SCL method integrates a context-aware structured classifier population generation (CA-SCPG) module with a knowledge-aware structured classifier population supervision (KA-SCPS) module.
• CA-SCPG utilizes a high-level context-aware dense connection (HLCA-DC) mechanism within a recurrent-dense-siamese decoder (RDS-Decoder) to enhance feature propagation.
• KA-SCPS employs a generalized weighted cross-entropy loss and a novel knowledge-aware Dice loss (KA-DL) for harmonized classifier supervision.

Main Results:

• Experiments on PET/CT volumes for malignant melanoma, lymphoma, and lung cancer demonstrate the effectiveness of the SCL method.
• SCL significantly improved segmentation performance compared to state-of-the-art methods and baseline approaches.
• The knowledge-aware Dice loss effectively harmonized the learning process across the classifier population.

Conclusions:

• The proposed SCL method offers a superior approach to collaborative learning for medical image segmentation.
• SCL effectively enhances feature propagation and ensures synchronized learning among classifiers, leading to improved segmentation accuracy.
• This method shows significant promise for clinical applications in cancer detection and characterization using medical imaging.