Deep learning framework integrating 18F-FDG PET/CT and clinical data for early prediction of chemotherapy response in advanced breast cancer

This study aimed to evaluate the potential of 18 F-Fluorodeoxyglucose Positron Emission Tomography/Computed Tomography ( 18 F-FDG PET/CT) parameters, including advanced texture features, alongside clinical data (CD), to predict pathological complete response (pCR) after the first course of neoadjuvant chemotherapy (NAC) in breast cancer (BC) patients. A cohort of 204 BC patients underwent 18 F-FDG PET/CT imaging before (baseline, B) and after (follow-up, F) the first course of NAC. The nnUNet deep learning model segmented breast cancer lesions, and radiomic/metabolic features were extracted from delayed PET/CT images. CD was integrated with imaging features to predict pCR. Machine learning models were trained on B + F scans and CD, using either separated or combined Luminal, HER2, and Triple Negative (TN) subtype models. We implemented data augmentation techniques to solve the data imbalance problem, and the recursive feature elimination method was used to rank the top contributing features to the model’s prediction. An external dataset from the QIN Breast database was harmonized and used to test the generalization of the developed model. The model trained on all breast cancer subtypes achieved the highest performance, with a balanced accuracy of 0 . 85 ± 0 . 06 , significantly outperforming subtype-restricted models. Specifically, it outperformed the combined HER2+TN model (BAcc = 0 . 78 ± 0 . 11 , p a d j = 0 . 014 ), the HER2 model (BAcc = 0 . 73 ± 0 . 09 , p a d j = 0 . 011 ), and the TN model (BAcc = 0 . 68 ± 0 . 11 , p a d j = 0 . 008 ). Feature selection demonstrated that using a reduced set (75% = 3/4 of features) optimized performance (AUC: 0 . 85 ± 0 . 06 ), significantly outperforming both the full feature set (100%; AUC: 0 . 80 ± 0 . 07 , p a d j = 0 . 013 ) and the minimal set (50% = 1/2 of features; AUC: 0 . 77 ± 0 . 03 , p a d j = 0 . 011 ). For the HER2 model, PET/CT (B + F) + CD achieved a BAcc of 0 . 73 ± 0 . 09 and an F1-score of 0 . 67 ± 0 . 04 , significantly outperforming the TN model using PET/CT (B + F) + CD (BAcc: 0 . 68 ± 0 . 11 , p a d j = 0 . 012 ). Baseline PET/CT features alone also demonstrated superior predictive performance (BAcc: 0 . 71 ± 0 . 06 ) compared to follow-up scans (BAcc: 0 . 66 ± 0 . 03 , p a d j = 0 . 017 ). On the external QIN-BREAST dataset, the PET/CT model using combined T1 + T2 data achieved the highest performance (BAcc = 0 . 74 ± 0 . 05 ), significantly outperforming the PET-only model (BAcc = 0 . 70 ± 0 . 08 , p a d j = 0 . 014 ) and the CT-only model (BAcc = 0 . 57 ± 0 . 07 , p a d j = 0 . 003 ). In conclusion, integrating baseline and follow-up 18 F-FDG PET/CT with CD significantly enhances pCR prediction in BC patients undergoing NAC.