Classification of pleural effusions using deep learning visual models: contrastive-loss.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pulmonology

Background:

• Blood and fluid analysis are crucial for diagnosing pleural effusion etiology.
• Current research often focuses on disease presence rather than etiological classification.
• Advanced computational methods are needed to improve diagnostic accuracy.

Purpose of the Study:

• To classify the etiology of pleural effusion using deep learning models.
• To evaluate the performance of a contrastive-loss deep learning model against other models.
• To visualize the model's decision-making process for clinical insights.

Main Methods:

• Analysis of patient data with pleural effusion undergoing thoracentesis (2009-2019).
• Comparison of five deep learning models, with a focus on a contrastive-loss model.
• Evaluation using metrics like top-1/top-2 accuracy and AUROC (Area Under the Receiver Operating Characteristic Curve).
• Visualization of the model's embedding space using UMAP and t-SNE.

Main Results:

• The contrastive-loss model achieved the highest accuracy on the extra-validation set.
• Validation set performance: 81.7% accuracy and 0.942 micro-AUROC.
• Extra-validation set performance: 66.2% accuracy and 0.867 micro-AUROC.
• Embedding space visualization successfully distinguished typical and atypical effusion cases.

Conclusions:

• Deep learning, specifically the contrastive-loss model, is effective for classifying pleural effusion etiology.
• Model visualization offers valuable insights for clinicians in differentiating disease types.
• This approach enhances diagnostic capabilities for pleural effusion.