Long-tailed medical diagnosis with relation-aware representation learning and iterative classifier calibration

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Computer-aided diagnosis (CAD) tools alleviate clinician workload but struggle with imbalanced datasets, often misdiagnosing rare diseases.
• Existing methods for long-tailed problems in medical imaging face challenges with biased learning and poor classifier calibration due to limited rare-class samples.

Purpose of the Study:

• To introduce a novel Long-tailed Medical Diagnosis (LMD) framework for balanced medical image classification.
• To enhance the performance of computer-aided diagnosis systems on datasets with significant class imbalance.

Main Methods:

• Developed a Relation-aware Representation Learning (RRL) scheme to improve feature extraction by leveraging data augmentation.
• Proposed an Iterative Classifier Calibration (ICC) scheme using Expectation-Maximization to generate balanced virtual features and refine classifiers.
• Implemented a two-stage approach focusing on representation learning and classifier calibration for imbalanced medical data.

Main Results:

• The LMD framework demonstrated significant improvements over state-of-the-art methods on three public long-tailed medical datasets.
• The RRL and ICC schemes effectively addressed biased representation learning and insufficient classifier calibration.
• Achieved superior performance in classifying minority (rare) disease categories within imbalanced medical image datasets.

Conclusions:

• The proposed LMD framework offers a robust solution for balanced medical image classification in the presence of long-tailed data distributions.
• The combination of RRL and ICC effectively mitigates performance disparities between majority and minority classes in computer-aided diagnosis.
• The framework holds promise for improving the reliability and fairness of diagnostic AI in clinical practice.