Enhancing Multi-Label Chest X-Ray Classification Using an Improved Ranking Loss.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Chest X-ray (CXR) analysis for thoracic diseases presents a multi-label classification challenge due to co-occurring pathologies.
• Class imbalance in CXR datasets, where some diseases are rare, complicates accurate model training.
• Training deep learning models from scratch requires massive datasets, often unavailable for specific medical tasks.

Purpose of the Study:

• To develop an effective method for multi-label classification of thoracic diseases in CXR images.
• To address the challenge of class imbalance in CXR datasets.
• To improve the performance of deep learning models in thoracic disease detection.

Main Methods:

• Utilized transfer learning by fine-tuning a pretrained DenseNet121 model on the NIH Chest X-ray14 dataset.
• Proposed a novel rank-based loss function, focal ZLPR (FZLPR), inspired by focal loss to address class imbalance.
• Incorporated a temperature parameter in FZLPR to emphasize hard-to-classify instances of rare diseases.

Main Results:

• The proposed FZLPR loss function outperformed standard loss functions like binary cross entropy (BCE) and focal loss.
• Models trained with FZLPR demonstrated superior performance on the NIH Chest X-ray14 dataset.
• Achieved a competitive average AUC of 80.96% using test-time augmentations with the FZLPR-trained model.

Conclusions:

• The focal ZLPR loss function is a promising approach for improving multi-label classification of thoracic diseases in CXR images.
• The methodology effectively handles class imbalance, leading to more accurate detection of both common and rare diseases.
• The developed model shows competitive performance, offering potential for enhanced diagnostic support in clinical settings.