Combo loss: Handling input and output imbalance in multi-organ segmentation

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Accurate multi-organ segmentation is vital for computer-aided diagnosis, surgery, and therapy planning.
• Deep learning models have shown promise but struggle with class imbalance in medical imaging data.
• Input (varying organ sizes, small foregrounds) and output (false positives/negatives) imbalances pose significant challenges.

Purpose of the Study:

• To develop and evaluate a novel loss function for deep multi-organ segmentation that addresses input and output data imbalance.
• To improve the robustness and accuracy of medical image segmentation models.

Main Methods:

• Introduced a curriculum learning-based loss function integrating Dice similarity coefficient and cross-entropy terms.
• The Dice coefficient prevents local minima, while cross-entropy penalizes false positives/negatives.
• Evaluated on diverse datasets: whole-body PET (5 organs), MRI prostate scans, and ultrasound (left ventricle).

Main Results:

• The proposed loss function, when used with a simple network, outperformed state-of-the-art methods on multiple datasets.
• Integrating the novel loss function improved the performance of competing methods.
• Demonstrated effectiveness across different modalities and segmentation tasks (multi-organ and single-organ).

Conclusions:

• The proposed integrative loss function is effective in handling data imbalance for multi-organ medical image segmentation.
• This approach offers a significant improvement over existing methods and enhances the performance of other models.
• The method shows broad applicability across various medical imaging modalities and segmentation challenges.