Predicting PIK3CA mutation in breast cancer with machine learning based multimodal image radiomics

Area of Science:

• Radiology
• Oncology
• Machine Learning

Background:

• Gene status is crucial for determining targeted therapy in breast cancer (BC).
• Early prediction of PIK3CA mutations can guide treatment decisions.
• Radiomics analysis of mammography (MMG) and ultrasound (US) images offers potential for non-invasive mutation prediction.

Purpose of the Study:

• To explore machine learning (ML) approaches for predicting PIK3CA mutations in BC.
• To evaluate the efficacy of MMG and US radiomics, individually and combined, for PIK3CA mutation prediction.
• To develop a hybrid model integrating clinicopathological and radiomics features for enhanced predictive performance.

Main Methods:

• 186 primary BC patients with PIK3CA gene testing and pretreatment MMG/US images were analyzed.
• Radiomics features were extracted from MMG and US images using 3D Slicer and PyRadiomics.
• Logistic Regression (LR), Adaptive Boosting (AdaBoost), and Naive Bayes (NB) ML algorithms were employed to build predictive models.

Main Results:

• PIK3CA mutations were identified in 31.7% of BC patients.
• Estrogen receptor (ER) and progesterone receptor (PR) positivity were associated with PIK3CA mutation.
• The hybrid clinicopathological-US-MMG-LR model achieved the highest predictive performance (AUC, 0.899), outperforming models based on single imaging modalities or other ML algorithms.

Conclusions:

• Ultrasound (US)-based radiomics demonstrated superior predictive ability compared to mammography (MMG)-based radiomics.
• The hybrid model integrating clinicopathological data with US and MMG radiomics significantly improved prediction accuracy.
• This advanced predictive model can aid in developing tailored therapeutic strategies for breast cancer patients.