Integrating Deep Learning of Low-Dose CT Imaging With Clinical Data for Lung Cancer Risk Prediction.

Area of Science:

• Radiology and Imaging
• Oncology
• Artificial Intelligence in Medicine

Background:

• Low-dose computed tomography (LDCT) screening is crucial for reducing lung cancer mortality.
• Segmentation-free deep learning (DL) models like Sybil offer potential screening efficiency improvements but need further validation and enhancement.

Purpose of the Study:

• To investigate if integrating DL with LDCT scans and clinical data can enhance lung cancer risk prediction.
• To evaluate the performance of a novel integrated model (Sybil-Epi) compared to existing DL models.

Main Methods:

• Retrospective analysis of 52,482 LDCT scans from 22,469 participants across four screening programs (2002-2021).
• Trained and externally validated DL models (Sybil and Sybil-Epi) for lung cancer risk prediction up to 6 years post-scan.
• Calculated Area Under the Receiver Operating Characteristic Curve (AUC) stratified by nodule presence and size; evaluated clinical-epidemiological factors.

Main Results:

• Sybil demonstrated AUCs from 0.93 (year 1) to 0.79 (year 6), with suboptimal performance for absent or small nodules (AUC=0.64, 0.61).
• The integrated Sybil-Epi model achieved a higher AUC (0.83) compared to Sybil (0.80) at year 6.
• Sybil-Epi showed significantly improved AUC (0.76) versus Sybil (0.64) when nodules were absent.

Conclusions:

• While Sybil is effective for short-term risk, its accuracy decreases when nodules are not detected.
• The Sybil-Epi model, incorporating DL with clinical-epidemiological data, significantly enhances lung cancer risk prediction accuracy, particularly in nodule-absent cases.