A Self-supervised Learning-Based Fine-Grained Classification Model for Distinguishing Malignant From Benign...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Diagnosing subcentimeter solid pulmonary nodules (SSPNs) is clinically challenging.
• Deep learning (DL) shows promise in improving pulmonary nodule classification over conventional methods.

Purpose of the Study:

• To develop and validate a DL model for differentiating malignant and benign SSPNs using CT images.
• To assess the model's performance in both internal and external validation cohorts.

Main Methods:

• A retrospective study utilizing CT images from 1276 patients with SSPNs.
• Development of a self-supervised pre-training-based fine-grained network for malignancy prediction.
• Model validation on internal (316 SSPNs) and external (202 SSPNs) datasets.

Main Results:

• The DL model achieved high performance: internal AUC 0.964, accuracy 0.934; external AUC 0.945, accuracy 0.911.
• Excellent sensitivity (internal 0.965, external 0.977) and specificity (internal 0.908, external 0.860) were observed.
• The model demonstrated robust performance across both datasets.

Conclusions:

• The developed deep learning model is robust and effective for predicting SSPN malignancy.
• This tool has the potential to optimize clinical management and improve patient outcomes in lung nodule diagnosis.