179P - Multimodal deep learning integrating MRI and molecular profiles for predicting outcomes in triple-negative breast cancer

Background

Despite advancements in high-throughput technologies and data-driven methods enhancing our understanding of cancer outcomes, accurately predicting the clinical behavior of triple-negative breast cancer (TNBC) remains challenging. This study aims to develop a predictive system for TNBC prognosis and immune subtypes, including Basal-like, Mesenchymal, Immune-activated, and Luminal Androgen Receptor (LAR), by integrating medical imaging and molecular profiles without requiring regions of interest (ROI).

Methods

We utilized MRI image and molecular profile data from The Duke-Breast-Cancer-MRI project and Gangnam Severance Hospital as independent datasets. Our approach includes a unimodal prediction model using single, ROI-free MRI images, and a multimodal model combining MRI and molecular data in TNBC. We employed a custom ResNet152 convolutional neural network (CNN) for MRI image processing and a fully-connected neural network (FNN) for molecular data analysis.

Results

In predicting the 3-year disease-free survival (DFS) rates of TNBC across multiple cohorts, the ResNet152-based unimodal model demonstrated notable predictive accuracy (95.4%), while the multimodal model, incorporating molecular profiles like TP53 and FOXM1 transcripts, showed enhanced performance (96.8% accuracy). The models exhibited strong prognostic capabilities (log-rank tests, p < 0.05) and clinical utility (multivariate Cox regression model, hazard ratio = 1.51, 95% CI = 1.07-2.22, p = 0.01) in TNBC prognosis. In classifying TNBC immune types, the unimodal model achieved 96.7% accuracy, and the multimodal model showed further improvement with 97.3% accuracy, underscoring its potential in guiding immunotherapy based on immune checkpoint inhibitors (ICIs).

Conclusions

Our ResNet152-based multimodal prediction model demonstrated significant prognostic and therapeutic predictive value for ICI-based treatments in TNBC. This tool enhances the understanding of TNBC prognosis and treatment efficacy, potentially improving clinical practices and complementing existing diagnostic approaches.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

This research was supported by National Research Foundation of Korea (NRF) grants, funded by the Korean government (No. 2021M3H9A209695312, and 2022R1A2C200848011) and a grant from the Korea Research Institute of Bioscience and Biotechnology (KRIBB) Research Initiative Program (KGM5192423).

Disclosure

All authors have declared no conflicts of interest.