Identifying Solitary Granulomatous Nodules from Solid Lung Adenocarcinoma: Exploring Robust Image Features with...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate preoperative differentiation of solitary pulmonary nodules is crucial for patient management.
• Lung adenocarcinoma (LAC) and lung granulomatous nodules (LGNs) present similar imaging features, posing diagnostic challenges.

Purpose of the Study:

• To identify optimal source domain data for robust feature extraction in cross-domain transfer learning.
• To develop and validate a radiomics model for preoperative distinction between LGN and LAC in solitary pulmonary nodules (SPSNs).

Main Methods:

• Retrospective analysis of 841 patients with SPSNs from five centers.
• Adaptive cross-domain transfer learning utilized to construct and compare transfer learning signatures (TLS).
• A cross-domain transfer learning radiomics model (TLRM) integrated best TLS, clinical factors, and CT findings, validated across multicenter cohorts.

Main Results:

• Transfer learning signatures derived from lung whole slide images (TLS-LW) demonstrated superior performance (AUC 0.8228-0.8984) and minimal Wasserstein distance.
• The final TLRM, incorporating TLS-LW, age, spiculated sign, and lobulated shape, achieved high diagnostic accuracy (AUC 0.9074-0.9442) across validation cohorts.
• Decision curve analysis and integrated discrimination improvement confirmed TLRM's superior performance over other models.

Conclusions:

• The developed TLRM effectively assists physicians in the preoperative differentiation of LGN from LAC in SPSNs.
• Cross-domain transfer learning, particularly with lung whole slide images as source data, enables robust feature extraction for improved diagnostic performance.