PCCT: Progressive Class-Center Triplet Loss for Imbalanced Medical Image Classification

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Class imbalance in medical imaging datasets poses a significant challenge for diagnosing rare diseases.
• Accurate diagnosis of rare diseases is hindered by limited data for these conditions.

Purpose of the Study:

• To introduce a novel two-stage Progressive Class-Center Triplet (PCCT) framework to address class imbalance in medical image diagnosis.
• To improve the classification performance for rare diseases using imbalanced datasets.

Main Methods:

• The PCCT framework employs a two-stage approach: a class-balanced triplet loss in the first stage and a class-center involved triplet strategy in the second stage.
• The first stage ensures equal sampling per class to mitigate imbalance, while the second stage promotes compact class representations by using class centers.
• The framework's generalization is demonstrated through extensions to pair-wise ranking and quadruplet losses.

Main Results:

• The PCCT framework achieved state-of-the-art performance on four challenging imbalanced datasets (Skin7, Skin 198, ChestXray-COVID, Kaggle EyePACs).
• Mean F1 scores over all classes were 86.20, 65.20, 91.32, and 87.18 respectively.
• For rare classes, the PCCT framework achieved F1 scores of 81.40, 63.87, 82.62, and 79.09, outperforming existing methods.

Conclusions:

• The proposed PCCT framework is highly effective for medical image classification with imbalanced training data.
• PCCT significantly enhances diagnostic accuracy for rare diseases, offering a robust solution to the class imbalance problem.
• The class-center involved loss strategy demonstrates broad applicability and training stability.