Multi-task learning for the segmentation of organs at risk with label dependence

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare

Background:

• Accurate segmentation of organs at risk in CT images is vital for diagnosis.
• Current segmentation methods face challenges in precision and efficiency.

Purpose of the Study:

• To develop an advanced automatic segmentation method for organs at risk using multi-task learning (MTL).
• To improve the accuracy and reliability of organ segmentation in CT scans.

Main Methods:

• Implemented an encoder-decoder network trained on parallel segmentation and multi-label classification tasks.
• Introduced a weighted mean cross entropy loss function incorporating global conditional probabilities for enhanced classification.
• Optimized a false positive filtering (FPF) algorithm with a dynamic threshold selection (DTS) strategy to maintain true positive rates.

Main Results:

• The proposed MTL approach significantly improved segmentation performance compared to basic encoder-decoder networks.
• The weighted loss function and DTS strategy boosted classification accuracy and filtering effectiveness.
• Validation on public (SegTHOR) and private datasets confirmed the method's efficacy without increasing computational time.

Conclusions:

• The developed MTL framework offers a robust and efficient solution for automatic organ segmentation in medical imaging.
• This technique holds potential for enhancing diagnostic accuracy and treatment planning in clinical settings.