F-FDG PET/CT radiomic analysis and artificial intelligence to predict pathological complete response after neoadjuvant chemotherapy in breast cancer patients.

Purpose: Build machine learning (ML) models able to predict pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) in breast cancer (BC) patients based on conventional and radiomic signatures extracted from baseline [F]FDG PET/CT.

Material And Methods: Primary tumor and the most significant lymph node metastasis were manually segmented in baseline [F]FDG PET/CT of 52 newly diagnosed BC patients. Clinical parameters, NAC and conventional semiquantitative PET parameters were collected. The standard of reference considered was surgical pCR after NAC (ypT0;ypN0). Eight-hundred-fifty-four radiomic features (RFts) were extracted from both PET and CT datasets, according to IBSI; robust RFTs were selected. The cohort was split in training (70%) and validation (30%) sets. Four ML Models (Clinical Model, CT Model, PET Model_T and PET Model_T + N) each one with 3 learners (Random Forest (RF), Neural Network and Stochastic Gradient Descendent) were trained and tested using RFts and clinical signatures. PET Models were built considering robust RFTs extracted from either primary tumor alone (PET Model_T) or also including the reference lymph node (PET Model_T + N).

Results: 72 pathological uptakes (52 primary BC and 20 lymph node metastasis) at [F]FDG PET/CT were segmented. pCR occurred in 44.2% cases. Twelve, 46 and 141 robust RFts were selected from CT Model, PET Model_T and PET Model_T + N, respectively. PET Models showed better performance than CT and Clinical Models. The best performances were obtained by the RF algorithm of the PET Model_T + N (AUC = 0.83;CA = 0.74;TP = 78%;TN = 72%).

Conclusion: ML models trained on PET/CT radiomic features extracted from primary BC and lymph node metastasis could concur in the prediction of pCR after NAC and improve BC management.