Deep Learning Model for Diagnosing and Classifying Subtypes of Chronic Pulmonary Aspergillosis in Chest CT.

Abstract

Background: Diagnosing chronic pulmonary aspergillosis (CPA) and its subtypes is essential for treatment and prognosis. In clinical practice, inexperienced doctors may overlook the presence of CPA due to overreliance on radiological results. Applying deep learning technology enhances multi-classification model performance.

Objective: To explore whether artificial intelligence generation technology and semisupervised learning can enhance model performance in CPA diagnosis and accurately classify CPA subtypes using small-sample datasets with skewed distributions and multiclass features.

Methods: This study developed a multi-classification model for CPA diagnosis and subtype differentiation using a multi-centre CT dataset. We augmented the small, skewed dataset with generation models and trained the deep learning model through a semi-supervised algorithm. Overfitting and poor validation generalisation issues were addressed with the internal dataset. The model, trained with different strategies, was evaluated on multiple internal and external test sets, measuring diagnostic performance via sensitivity, accuracy, F1 score, Matthews correlation coefficient, CK score and overall accuracy.

Results: A total of 39,387 chest CT images from 660 patients were split into training, validation and internal test sets. Additionally, 3337 chest CT images from 11 patients formed external test set 1, while 120 images from other studies made up external test set 2. The optimal model successfully diagnosed six CPA patients hidden in external test set 1 and classified their subtypes. In external test set 2, it achieved an ACC of 91% and an AUC of 0.92.

Conclusion: Using synthetic data and semi-supervised learning improved deep learning performance in diagnosing and classifying chronic pulmonary aspergillosis on chest CT images.