Long-Tailed Classification of Thorax Diseases on Chest X-Ray: A New Benchmark Study

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Medical image datasets often exhibit a 'long-tailed' distribution, with common diseases overrepresented and rare diseases underrepresented.
• Standard machine learning models struggle with long-tailed data, leading to bias towards frequent classes and poor performance on rare conditions.
• Accurate classification of rare thorax diseases from chest X-rays is crucial for effective patient diagnosis and treatment.

Purpose of the Study:

• To establish a comprehensive benchmark for evaluating long-tailed learning methods in medical image classification, specifically for thorax diseases on chest X-rays.
• To facilitate research and development of algorithms capable of learning from naturally distributed medical imaging data, encompassing both common and rare conditions.
• To optimize classification accuracy for both frequent ('head') and infrequent ('tail') thorax disease classes.

Main Methods:

• Introduction of a novel long-tailed chest X-ray benchmark dataset for thorax disease classification.
• Creation of two datasets for 19- and 20-way classification, featuring a significant imbalance in labeled training images (ranging from 7 to over 53,000 per class).
• Evaluation of standard and state-of-the-art long-tailed learning methods on the benchmark dataset.

Main Results:

• Analysis of the performance of various long-tailed learning methods on the challenging chest X-ray benchmark.
• Identification of key aspects of these methods that prove most beneficial for long-tailed medical image classification.
• Insights derived from the evaluation to guide future algorithm design for improved medical image analysis.

Conclusions:

• The developed benchmark provides a crucial resource for advancing long-tailed learning in medical imaging.
• Understanding the efficacy of different methods offers valuable guidance for creating more robust and equitable AI diagnostic tools.
• Future algorithm development should focus on addressing the specific challenges posed by imbalanced medical datasets to improve rare disease detection.