The five surrogate subtypes at the heart of the study are not academic labels. They are defined by the tumor’s expression of the estrogen receptor, the progesterone receptor, and HER2, the human epidermal growth factor receptor 2, and they map directly onto the most consequential treatment decisions in oncology. Hormone receptor-positive tumors are candidates for endocrine therapy, HER2-positive disease can be attacked with targeted antibodies such as trastuzumab, and the triple-negative subtype, which lacks all three receptors, is the most aggressive form and demands chemotherapy-based regimens. Following the 2024 edition of the breast cancer diagnosis and treatment guidelines issued by the China Anti-Cancer Association, the team categorized each of the 506 tumors into one of these five IHC-defined surrogate classes, creating the ground truth against which every model was judged.

The imaging substrate for the prediction is ¹⁸F-FDG itself, a radioactive glucose analog that cancer cells avidly take up because of their accelerated metabolism. In a PET/CT scan, the concentration of this tracer paints a functional portrait of the tumor: how hot it is, how its intensity is distributed, how its metabolic volume relates to its anatomical size. Radiologists have long used qualitative versions of these cues, but the Tianjin team wanted to know whether deep neural networks could extract far richer, subtler metabolic fingerprints that correlate with receptor status, fingerprints invisible to the human eye.

The technical architecture of the study reflects a careful, comparative design. The researchers constructed four predictive models from clinical and imaging data collected between January 2010 and December 2024. Two were pure deep learning models that consumed PET images directly: EfficientNet-B0, a lightweight convolutional network known for computational efficiency, and SEResNext50, a heavier architecture that combines residual connections with squeeze-and-excitation blocks that let the network recalibrate the importance of different feature channels. The third was a decision fusion model that merged the probability outputs of the two deep networks, and the fourth was a machine learning model built on LightGBM, a gradient boosting framework that integrated image-derived deep features with structured clinical variables.

The performance hierarchy that emerged was instructive. EfficientNet-B0, despite its popularity, showed suboptimal classification performance and limited stability when asked to distinguish the five subtypes from PET images alone. SEResNext50 fared better, demonstrating moderate utility as a foundation for ensemble learning. The decision fusion model, which combined the two networks’ probability outputs, effectively compensated for the weaknesses of each single architecture, yielding improved predictive accuracy and stability. This pattern illustrates a principle well known in machine learning but rarely demonstrated so cleanly in a clinical imaging context: ensembles of diverse models are more robust than any of their components because their individual errors tend to cancel rather than compound.

But the star of the study was the LightGBM model. By taking features that the deep networks had learned from the PET images and feeding them, alongside clinical indicators, into a gradient boosting machine, the team achieved an AUC of 0.95, with a 95 percent confidence interval of 0.94 to 0.96, on the validation set, and 0.86, with a confidence interval of 0.84 to 0.88, on the held-out test set. The gap between validation and test performance is a sobering reminder of the challenge of generalization in medical AI, but the test-set figure still represents strong discriminative power for a five-class problem. The researchers assessed the models comprehensively with receiver operating characteristic curves, precision-recall curves, confusion matrices, and classification error plots, and these analyses confirmed that the LightGBM model possessed the best classification capability, reasonable internal robustness, and good overall stability.

What makes this result scientifically interesting is the fusion strategy itself. PET images alone carry metabolic information, but they do not know the patient’s age, tumor size, or histological type. Clinical features alone carry demographic and anatomical context, but they cannot see the tumor’s metabolic behavior. By extracting deep features from the images and combining them with structured clinical variables in a single gradient boosting model, the researchers created a hybrid representation that outperformed both pure imaging approaches and pure feature-based approaches. The LightGBM framework, which builds an ensemble of decision trees in a sequential manner with each tree correcting the errors of its predecessors, proved particularly well suited to this heterogeneous feature space, where image-derived embeddings and clinical measurements coexist at very different scales and distributions.

The clinical implications are substantial. Molecular subtyping currently requires tissue, and tissue requires a biopsy, which is invasive, costly, and subject to sampling error, since a needle may capture only a fraction of a heterogeneous tumor. A validated AI model that reads a PET scan could offer a non-invasive preliminary estimate of subtype, potentially guiding the urgency and design of subsequent biopsies, flagging discordances between imaging predictions and pathology results, and providing an objective reference for individualized treatment planning. In settings where PET/CT is already part of the staging workup, the marginal cost of such a prediction approaches zero, since the scan has already been acquired.

The study’s limitations are those inherent to its retrospective, single-center design. All 506 patients were imaged at one institution over fourteen years, and the models have not yet been tested on external cohorts from other hospitals with different scanners, protocols, and patient populations. The drop in AUC from 0.95 on validation to 0.86 on the test set suggests that some degree of overfitting to the development data occurred, and the authors themselves note that the models require further validation before clinical deployment. The five-class problem is also harder than binary classification, and performance on individual subtypes, particularly the rarer classes, may vary. Still, the comprehensive evaluation with ROC curves, PR curves, confusion matrices, and error plots across all four models provides a transparent picture of where each approach succeeds and fails.

Looking forward, the fusion paradigm demonstrated here, in which deep learning extracts imaging features and gradient boosting integrates them with clinical data, could extend beyond breast cancer to other tumors where receptor status drives therapy, and beyond PET to other imaging modalities. As the authors conclude, ¹⁸F-FDG PET integrated with deep learning-based machine learning models provides significant auxiliary diagnostic value for the non-invasive molecular subtyping of breast cancer. If external validation confirms these findings, the humble PET scan, already a staple of cancer staging, could evolve into a molecular diagnostic instrument, offering patients a faster, less invasive path to the subtype information that shapes their treatment.

Subject of Research: breast cancer

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

Article References: Fu, Q., Bian, H., Ma, W., Li, X., Zhu, Y., Sun, R., Wang, J., Chen, W., Liu, J., Xu, W., & Dai, D. (2026). Application of a deep learning-based ¹⁸F-FDG PET and clinical-feature fusion model in IHC-defined surrogate subtyping of breast cancer. BMC Cancer. https://doi.org/10.1186/s12885-026-16987-z

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16987-z

Keywords: breast cancer, molecular subtyping, deep learning, PET imaging, machine learning, LightGBM, ¹⁸F-FDG PET, Application, deep, learning-based, F-FDG, clinical-feature