Harmony Loss for Unbalanced Prediction | JoVE Visualize

Area of Science:

Medical Image Analysis
Deep Learning
Machine Learning

Background:

Deep learning models in medical imaging often struggle with unbalanced datasets, leading to poor performance on underrepresented categories.
The Area Under the Precision-Recall Curve (AUC_PR) is a metric sensitive to class imbalance, but its discrete nature poses challenges for deep learning optimization.
Existing loss functions may not effectively reconcile precision and recall across all categories, especially minority classes.

Purpose of the Study:

To develop a novel loss function, Harmony loss, that mitigates the impact of unbalanced datasets in medical image analysis.
To ensure continuous differentiability and gradient existence for optimization by approximating the AUC_PR calculation.
To improve the convergence speed and stability of deep learning models trained on imbalanced medical imaging data.

Main Methods:

Constructed a Harmony loss function leveraging the sensitivity of AUC_PR to different sample categories.
Employed the Logistic function to approximate the logical function within AUC_PR, ensuring continuous differentiability.
Implemented a method of setting discrete classification thresholds to approximate AUC_PR, enhancing computational efficiency and optimization speed.
Validated the Harmony loss function across 3D reconstruction, 2D segmentation, and unbalanced classification tasks.

Main Results:

The Harmony loss function demonstrated fast convergence and high stability during model training.
Achieved state-of-the-art performance on four unbalanced datasets, outperforming existing methods.
Successfully improved the model's ability to recognize categories with fewer samples without compromising overall stability.
Showcased compatibility with existing loss functions and suitability for common deep learning architectures.

Conclusions:

The Harmony loss function is an effective solution for addressing class imbalance in medical image analysis.
Its design ensures stable gradients and computational efficiency, making it practical for deep learning applications.
The Harmony loss offers a versatile and high-performing alternative for improving deep learning model robustness on imbalanced medical data.