Clinical Study of Imaging Genomics Based on Machine Learning for Breast Cancer Molecular Typing and Risk Prediction (BCIG)

Research design

1. Research on the molecular typing of breast cancer based on imaging features
2. Establish a Luminal breast cancer recurrence risk prediction model
3. Establish HER2 targeted therapy sensitivity prediction model
4. Establish TNBC molecular subtype prediction model Research methods Research Object This study used a multi-center study to prospectively enroll breast cancer patients diagnosed with pathology. All enrolled patients had complete clinical data, including demographic characteristics (gender, age, menstrual status and fertility history), and pathological data (histopathological data). Staging, immunohistochemical status and FISH, genetic testing records the recurrence score and genotype), imaging data, complete treatment and follow-up (whether there is local recurrence and metastasis, and the time of diagnosis).

Magnetic resonance examination In order to maintain the comparability between the images and reduce the systematic errors, each center selects a fixed MR device for scanning. Among them, a. Oncology Hospital chose to scan images with 3.0T (Siemens Skyra) MR equipment. A special breast coil is used to add high-definition diffusion-weighted scanning and multi-b value diffusion-weighted scanning before the dynamic enhancement scan. Dynamically enhanced acquisition in 5 phases with a time resolution of 65s. b. Renji Hospital uses Netherlands Philips Achieva 3.0 T superconductor MR scanner, 4-channel dedicated breast phased array coil. Scanning sequences include T1WI, T2WI, T2WI fat suppression, DWI and DCE-MRI. The contrast agent was Gd-DTPA, with a dose of 0.1 mmol/kg, an injection rate of 2.0 mL/s, and an additional 20 mL of saline was added to the tube after injection. The T1WI scan was performed first, and 5 time phases were continuously scanned after the injection of contrast agent, and each time phase was separated by 61 s, for a total of 6 time phases. c. Chinese women and babies are scanned with 1.5T SIEMENS AERA MR equipment and special breast coils. Scanning sequence includes 5 phases of T1WI, T2WI fat suppression, DWI and dynamic enhancement scan, time resolution 71s.

Image processing Use software to make semi-automatic and automatic outlines of the tumor interest area, and make the outline of the tumor solid enhancement part, the entire tumor area and the surrounding edema zone in the transverse position. In order to accurately delineate the tumor, compare the T1 and T2 weighted and dynamically enhanced images, two imaging physicians are responsible, one is responsible for delineation and the other is reviewed, and the disputed area is determined after discussion by a third person. Create a dynamic enhanced tumor texture analysis program to automatically extract imaging omics features in the region of interest. Using a labeled data set, a computer-based automatic segmentation algorithm model based on machine learning is constructed to automatically extract regions of interest, and segmentation performance evaluation is performed on manually delineated labels.

Statistical analysis Perform statistical analysis on the obtained images and clinical data, extract image omics features and use machine learning algorithms to screen important features. Use statistical tools such as SPSS and R language. Paired t test (continuous variable) and chi-square test (discontinuous variable) were used to compare the clinical and imaging characteristics of patients with different prognosis; correlation analysis was used to evaluate the imaging histology characteristics and different pathological tissue grades, Correlation between lymph node metastasis and specific gene expression; use Kaplan-Meier survival curve to analyze the prognostic difference between patients with different imaging omics characteristics, and use log-rank method to test the difference; use cox survival model to compare clinical characteristics and imaging omics The characteristics and prognosis of patients (tumor-free survival, progression-free survival, overall survival) were analyzed by multiple factors. Further, deep learning algorithms can be used to automatically learn imaging omics features that may be related to molecular subtypes and prognosis to build prediction models.