Application of a deep learning-based ¹⁸F-FDG PET and clinical-feature fusion model in IHC-defined surrogate subtyping of breast cancer

This study aimed to investigate the diagnostic performance of ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) in conjunction with advanced artificial intelligence models for the molecular subtyping of breast cancer. Clinical and imaging data from 506 patients diagnosed with primary breast cancer who underwent ¹⁸F-FDG PET/CT at Tianjin Medical University Cancer Institute and Hospital between January 2010 and December 2024 were constructed retrospectively. Following the Guidelines and Specifications for Diagnosis and Treatment of Breast Cancer (2024 Edition) issued by the China Anti-Cancer Association, tumors were categorized into five IHC-defined surrogate subtypes based on immunohistochemistry (IHC) profiles. Four predictive models were developed and compared: two PET image-based deep learning models, one decision fusion model, and one feature-based machine learning model incorporating clinical and imaging features. Model performance was comprehensively assessed using receiver operating characteristic (ROC) curves, precision-recall (PR) curves, confusion matrices, and classification error plots. Four predictive models were established in this study. Among the single deep learning models, EfficientNet-B0 exhibited suboptimal classification performance and limited stability, whereas SEResNext50 demonstrated moderate utility as a foundation for ensemble learning. The decision fusion model, based on dual-model probability fusion, effectively mitigated the limitations inherent in single deep learning models, yielding improved predictive accuracy and stability. The light gradient boosting machine (LightGBM) learning model, constructed by integrating image-derived deep features with clinical features, exhibited the best overall performance and generalization ability, achieving an area under the curve (AUC) of 0.95 (95% CI: 0.94–0.96) on the validation set and 0.86 (95% CI: 0.84–0.88) on the test set. Results from ROC curves, PR curves, confusion matrices, and classification error plots confirmed that the LightGBM model possesses better classification capability, reasonable internal robustness, and good overall stability. ¹⁸F-FDG PET, when integrated with deep learning-based machine learning models, provides significant auxiliary diagnostic value for the non-invasive molecular subtyping of breast cancer. The LightGBM machine learning model, constructed by combining imaging features with clinical indicators, can further enhance subtyping prediction performance, providing an objective reference for the formulation of individualized clinical treatment plans and providing an objective reference for individualized clinical treatment planning in breast cancer. This study was a retrospective study, so it was free from registration.