Breast Cancer Recurrence Risk Determined by Deep Learning Model Trained on Histopathologic Slides

A deep learning model demonstrated the ability to predict breast cancer recurrence risk and possible benefit from the addition of chemotherapy based on histopathologic images rather than genomic testing in patients with hormone receptor–positive, HER2-negative breast cancer, according to findings presented during the inaugural European Society for Medical Oncology (ESMO) Artificial Intelligence (AI) & Digital Oncology Congress (Abstract 172MO). The model would allow for clinicians in developing countries to determine Oncotype DX^® breast cancer recurrence scores from hematoxylin-and-eosin images alone and potentially avoid overtreating low-risk patients. 

“This can be very impactful in places where chemotherapy decisions are based on clinical risk and especially significant for reduction of overtreatment [in patients with hormone receptor–positive, HER2-negative breast cancer],” said presenting author Gil Shamai, PhD, of Technion – Israel Institute of Technology, Haifa. 

Background

Genomic testing has improved clinical decision-making regarding the use of chemotherapy in patients with hormone receptor–positive, HER2-negative breast cancer. For example, the Oncotype DX recurrence score is one such gene expression assay that has been integrated into the National Comprehensive Cancer Network Clinical Practice Guidelines (NCCN Guidelines) for Breast Cancer, with a preferred recommendation. The Guidelines now recommend, as per the results of the TAILORx trial, that patients with a recurrence score of 26 or above receive chemotherapy with endocrine therapy, as they are the most likely to benefit from added chemotherapy.

However, Dr. Shamai noted that these genomic assays are very costly and time consuming, making them difficult to access in developing countries. As such, in developing countries, treatment decision-making is more dependent on clinical risk. In India, for example, about 85% of patients with hormone receptor–positive, HER2-negative breast cancer are given chemotherapy.

Dr. Shamai and his coinvestigators sought to predict patients’ Oncotype Dx score directly from histopathologic images using deep learning “because this approach could offer a much more affordable solution, and faster, and is especially useful in places where genomic testing is less accessible.”

Training the Model

The investigators built upon the GigaPath foundation model, which is pretrained on 171,189 hematoxylin-and-eosin slides. The slides were each segmented into tissue and background and then divided into small image tiles. From there, the model extracted image features from the tiles that were used with a transformer encoder and multiple-instance learning to label the slides and determine Oncotype DX recurrence scores with clinical variables. The model was then fine-tuned and validated on patients from the TAILORx trial (n = 2,407) as well as on six external cohorts across the world (n = 13,781).

Model Outcomes

In the TAILORx validation set, the survival stratification for distant recurrence–free survival was found to be similar between the AI-based recurrence score (hazard ratio [HR] = 2.88, 95% confidence interval [CI] = 1.73–4.79; P < .001) and the genomic Oncotype DX recurrence score (HR = 2.60, 95% CI = 1.56–4.33; P < .001). The stratification also appeared to be similar across different patient subgroups, such as premenopausal and postmenopausal patients.

The model was then assessed for how chemotherapy benefit related to the AI-based recurrence risk assessments. For postmenopausal patients, those with an AI-based recurrence score of between 11 and 26 did not show any benefit from added chemotherapy in terms of recurrence-free survival (HR = 0.95, 95% CI = 0.71–1.27; P = .739). In premenopausal patients with an AI-based score between 16 and 26, recurrence-free survival seemed to improve with chemotherapy treatment (HR = 0.55, 95% CI = 0.36–0.82; P < .01).

“This makes our model the first evidence-based predictive test in breast cancer based on digital pathology,” Dr. Shamai noted.

When validated against Australian, Israeli, and U.S. cohorts, the model showed high predictive generalizability, with areas under the curve ranging from 0.832 to 0.903. “This shows that this foundation model approach can be transferable to new data,” he said.

To assess the impact of the model in developing countries, the investigators measured the risk of patients from the TAILORx trial using the MINDACT criteria for clinical risk. The model reclassified several patients (5.4%) from low risk by MINDACT criteria to high risk by the AI model’s criteria, and 30.1% were reclassified from high to low risk, indicating that they may not benefit from chemotherapy.

Dr. Shamai noted that, going forward, the researchers are initiating a clinical trial in India to further validate this foundation model. 

DISCLOSURE: Dr. Shamai reported no conflicts of interest.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.