Bidirectional Weighted Cross-Entropy for Imbalanced Pressure Injury Classification

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate pressure injury classification from clinical images is difficult due to small, imbalanced datasets.
• Existing methods struggle with the nuances of misclassification costs in imbalanced data.

Purpose of the Study:

• To introduce Bidirectional Weighted Cross-Entropy (BWCE), a novel loss function for imbalanced image classification.
• To evaluate BWCE's effectiveness in reducing clinically undesirable misclassifications, particularly minority-to-majority errors.

Main Methods:

• Developed BWCE, a cost-sensitive loss function incorporating true-label and predicted class weights for misclassified samples.
• Evaluated BWCE on 853 pressure injury images using 3x10 repeated stratified cross-validation.
• Compared BWCE against standard loss functions (CE, Focal Loss) and other imbalance-aware methods (CB-CE, LDAM).

Main Results:

• BWCE-log demonstrated competitive performance in recall, F1-score, and AUC, improving recall over standard Cross-Entropy (CE).
• BWCE-log significantly reduced minority-to-majority misclassifications compared to CE and Focal Loss.
• While BWCE showed promise, other methods like CB-CE and LDAM excelled in specific minority-focused metrics.

Conclusions:

• BWCE presents a valuable direction-aware, cost-sensitive alternative for addressing class imbalance in medical image classification.
• The findings suggest BWCE is effective for modeling specific error directions at the loss-design level.
• BWCE is a complementary approach rather than a universally superior replacement for existing imbalance-aware objectives.