Batch-balanced focal loss: a hybrid solution to class imbalance in deep learning

Area of Science:

• Medical Imaging Analysis
• Machine Learning for Healthcare
• Computer Vision in Ophthalmology

Background:

• Class imbalance in medical datasets poses a significant challenge for developing accurate diagnostic models.
• Convolutional Neural Networks (CNNs) are powerful tools for image classification but struggle with imbalanced data.
• Existing methods for handling imbalanced data often fall short in complex medical imaging tasks.

Purpose of the Study:

• To validate the effectiveness of a novel approach, batch-balanced focal loss (BBFL), for enhancing CNN classification performance on imbalanced datasets.
• To assess BBFL's ability to improve diagnostic accuracy in medical image classification tasks.

Main Methods:

• BBFL was developed by combining batch-balancing and focal loss strategies to address class imbalance and hard-sample importance.
• The approach was validated on two imbalanced fundus image datasets: retinal nerve fiber layer defect (RNFLD) and glaucoma.
• BBFL was compared against random oversampling (ROS), cost-sensitive learning, and thresholding using state-of-the-art CNNs (InceptionV3, MobileNetV2).

Main Results:

• For binary RNFLD classification, BBFL achieved 93.0% accuracy, 84.7% F1-score, and 0.971 AUC, outperforming other methods.
• In multiclass glaucoma classification, BBFL achieved 79.7% accuracy and 69.6% average F1-score, surpassing ROS, cost-sensitive learning, and random undersampling.
• Visual assessments using confusion matrices, t-SNE plots, and GradCAM confirmed BBFL's superior performance.

Conclusions:

• The BBFL-based learning method significantly improves CNN model performance in both binary and multiclass disease classification tasks with imbalanced data.
• BBFL offers a robust solution for enhancing diagnostic accuracy in medical imaging where data imbalance is prevalent.
• This approach holds promise for developing more reliable AI-powered diagnostic tools in ophthalmology and other medical fields.