Development of an AI model for pneumothorax imaging: Dataset and model optimization strategies for real-world deployment

Abstract

Purpose

This study develops an AI-assisted pneumothorax diagnosis system using deep learning and chest X-ray images to enhance diagnostic efficiency and accuracy, reduce radiologists' workload, and provide timely treatment. The system addresses limitations of traditional methods, which rely on subjective interpretation and are vulnerable to fatigue or inexperience.

Methods

The DenseNet121 model was employed using a chest X-ray dataset from a medical center in northern Taiwan, with a total of 6888 images’ divided into training (64 %), validation (16 %), and testing (20 %) sets. Image preprocessing involved normalization, data augmentation (rotation, translation, scaling, brightness adjustment), and standardization. The model was trained using stochastic gradient descent with an initial learning rate of 0.0016 for 150 epochs. Performance evaluation included accuracy, sensitivity, specificity, and AUROC, integrating with the hospital's PACS for real-time analysis.

Results

Initial testing yielded AUROC values of 94.52 % and 97.21 % for pneumothorax and mild pneumothorax groups. However, when applied to 6888 clinical images, the AUROC dropped to 62.55 %, resulting in 4294 false positives. Adjusting the dataset split and retraining with 1000 false positive images improved the AUROC from 62.55 % to 85.53 %.

Conclusions

The AI model shows potential in pneumothorax detection, but performance is influenced by data diversity, image quality, and clinical complexity. The model struggles to identify key areas in complex cases, indicating a need for attention mechanisms or region proposal networks (RPN). Expanding the dataset, optimizing preprocessing, and training separate models for different image locations could enhance performance further.