Predicting PIK3CA mutation in breast cancer with machine learning based multimodal image radiomics.
Authors
Affiliations (5)
Affiliations (5)
- Department of Ultrasound, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, 2nd Ruijin Road 197, Shanghai, 200025, People's Republic of China.
- Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Pathology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Comprehensive Breast Health Center, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Ultrasound, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, 2nd Ruijin Road 197, Shanghai, 200025, People's Republic of China. [email protected].
Abstract
Gene status may help determine molecular targeted therapy for patients with breast cancer (BC). The present study aimed to explore machine learning (ML) approaches for the early prediction of PIK3CA mutation based on mammography (MMG) and ultrasound (US) radiomics. We analysed 186 primary BC patients (130 in the training set and 56 in the test set) who underwent PIK3CA gene testing and with pretreatment MMG and US images from January 2020 to December 2023 at our institution. PIK3CA mutation results were used as the reference standard. The MMG and US radiomics features were extracted using 3D slicer and PyRadiomics software. Three machine learning (ML) algorithms, namely, logistic regression (LR), adaptive boosting (AdaBoost), and Naive Bayes (NB) were applied to construct and validate the radiomics and combined clinicopathological-radiomics predictive models. Fifty-nine (31.7%) BC patients (41 in the training set and 18 in the test set) had PIK3CA mutations. Clinicopathological (C) factors including estrogen receptor (ER) and progesterone receptor (PR) positivity were associated with PIK3CA mutation. The AUC values of the MMG-based radiomics were 0.656 (NB), 0.700 (AdaBoost), and 0.718 (LR), while those of the US-based radiomics models were 0.732 (NB), 0.754 (AdaBoost) and 0.788 (LR) in the test set. The predictive ability of clinicopathological-radiomics C + US + LR model was greater (AUC, 0.827; 95% CI 0.706, 0.942) than that of the C + MMG + LR model (AUC, 0.740; 95% CI 0.611, 0.867; p < 0.05). The hybrid C + US + MMG + LR model achieved greater efficacy (AUC, 0.899; 95% CI 0.783, 0.984) than the C + US + MMG + AdaBoost (AUC, 0.860; 95% CI 0.739, 0.974), and C + US + MMG + NB (AUC, 0.838; 95% CI 0.721, 0.952; all p < 0.05) models in the test set. US-based radiomics showed better predictive ability than MMG-based radiomics model. The hybrid clinicopathological and US + MMG + LR model with improved performance can assist tailored therapeutic strategies.