ESTRO 2024 - Abstract Book

S5059

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

1516

Digital Poster

Deep Learning-Based Subtyping of Triple-Negative Breast Cancer

Ting Huang, Jiazhou Wang

Fudan University Shanghai Cancer Center, Radiation Therapy Center, Shanghai, China

Purpose/Objective:

Triple-negative breast cancer (TNBC) is a highly heterogeneous disease, Fudan University Shanghai Cancer Center has further subdivided the previously established transcriptome-based subtypes, including LAR, BLIS, MES, and IM. This molecular subtyping is beneficial for improved diagnostic precision and targeted therapies. However, existing methods like immunohistochemistry and genetic analysis take several days and the cost may range from hundreds to thousands of dollars. This study is the first to propose a model based on medical imaging, specifically mammographic views, for subtyping triple-negative breast cancer.

Material/Methods:

A dataset comprising mammographic images from 369 TNBC patients treated between January 2019 and July 2023 was divided into a training set (269 cases) and a validation set (100 cases). The study initially trained a BI-RADS classification model using 3542 cases Asian women specifically tailored for Asian race and applied transfer learning to classify TNBC subtypes, using a five-fold cross-validation approach. LASSO (Least Absolute Shrinkage and Selection Operator) feature selection was employed to enhance the model's performance. This process involved selecting the most relevant features for TNBC subtype classification. The selected features were used to improve the model's accuracy.

Results:

Without pre-training in BI-RADS, the model achieved an accuracy of 0.55 on the training set and 0.55 on the validation set for TNBC subtype classification. After applying transfer learning based on the BI-RADS model to train a TNBC subtype classification model, the accuracy reached 0.93 on the training set and 0.57 on the five-fold testing set. The model, combined with LASSO feature selection, achieved accuracy of 0.83 on the training set and 0.6824 on the validation set(as shown in the image) , with a model specificity of 0.74 and sensitivity of 0.623.