In-Silico Virtual Biopsy Platform To Personalize Breast Cancer Treatment

Author(s): Arjun Moorthy

Background: Recent studies exhibit preliminary data on the relationship of MRI based imaging phenotypes of breast tumors to breast cancer molecular and genomic characteristics. This study serves to explore relationships between MRI imaging and clinically relevant breast cancer characteristics with acceptable accuracies.

Methods: We analyzed 87 patients from the TCIA/TCGA (The Cancer Imaging Atlas/The Cancer Genome Atlas) open source dataset with invasive breast cancer and pre-operative MRI. LifeX open source software was used to extract radiomic features from MRI images. Machine learning based models based on the radiomic and imaging features were used to predict molecular subtype, recurrence score, novel miRNA correlations and biological pathways from the Hallmark GSEA dataset.

Results: Our models were able to use the radiomic analysis upon MRI images to predict the molecular subtype, risk of recurrence, miRNA expression, and genetic pathway expression. However, of these correlations the most accurate was the prediction of triple negative vs. non-triple negative cancers. The accuracy of the aforementioned correlations was around 92% (p-value = 0.02, 95% CI), while the other remaining correlations were around 69-73% accurate (95% CI), not high enough to be used in clinical practice, but a promising result that can be aided by larger datasets. Risk of recurrence was predicted with a 69-73% accuracy, and an imaging surrogate for miRNA 940 was identified.