5529. Machine Learning for the Classification of Molecular Subtype and Receptor Status of Breast Cancer Tumors in DCE-MRI
Authors * Denotes Presenting Author
  1. Michelle Le *; Penn State College of Medicine
  2. Gavin Jones; Penn State College of Medicine
  3. Sangam Kanekar; Penn State College of Medicine
  4. Alison Chetlen; Penn State College of Medicine
  5. Scott Hwang; Penn State College of Medicine
A role that artificial intelligence (AI) and machine learning (ML) can play in radiology is in the classification of MR images. The aim of this study is to use a commercially available machine learning tool to create a program capable of characterizing breast cancer tumors using lesions localized on DCE-MRI. The following characteristics will be investigated: ER status, PR status, HER2 status, and molecular subtype. Timely classification of tumor receptor status and subtype leads to faster delivery of appropriate therapy for patients with breast cancer.

Materials and Methods:
The AI tool was created using the cloud-based ML software Custom Vision to train models to classify a tumor’s receptor status and molecular subtype. The model was trained on a large library of well-annotated breast MR images of biopsy confirmed breast cancer. Using the coordinates of the localized breast cancer, multiple 2D slices of each tumor made up the training set. The contrast enhancement data were utilized by combining precontrast tumor slices with corresponding first and third postcontrast slices into one RGB image. The training data was augmented by horizontally flipping and adding variable noise to the tumor images so that each class had 1500 images. Four models were trained that classified tumors based on the following characteristics: ER+ or ER-, PR+ or PR-, HER2+ or HER2-, luminal A or luminal B, or HER2-enriched or triple negative.

The trained models are tested within the software and given scores based on model precision, recall, and mean average precision (AP). The ER model had a precision of 90.3%, and recall of 93.3%, and an AP of 97.0%. The PR model had a precision of 60.4%, and recall of 87.3%, and an AP of 76.0%. The HER2 model had a precision of 63.4%, and recall of 87.5%, and an AP of 76.9%. The molecular subtype model had a precision of 65.6%, and recall of 48.9%, and an AP of 64.0%.

Our study shows that there may be features in DCE-MRI of breast cancer that an ML software can detect allowing it to classify tumors based on receptor status. Performance varied based on the specific characteristic being trained, with ER status having the best overall results. The classification of molecular subtype had the weakest results. Since molecular subtype is based on receptor status, the determination of ER, PR, and HER2 status could be used to label the subtype. Based on tumor cell receptor profiles, breast cancer is divided into four molecular subtypes that guide treatment with one or a combination of either hormone therapy, monoclonal antibodies, or chemotherapy. Classically, there is a delay from breast biopsy to determination of molecular subtype. An accurate ML tool that can classify tumors would provide patients with treatment options earlier in their care.