Researchers at the Technion – Israel Institute of Technology have developed an artificial intelligence (AI) model designed to help guide whether to administer chemotherapy directly after tumor resection in breast cancer patients. The model, which uses high-resolution images from hematoxylin and eosin (H&E)-stained tumor samples collected at the time of diagnosis, predicts both the risk of cancer recurrence and the likelihood that a patient will benefit from chemotherapy. The research, published in The Lancet Oncology, identifies patterns and signals in the tumor that might otherwise escape detection.
“These are complex biological signals that the human eye cannot consistently quantify,” said first author Gil Shamai, PhD, a postdoctoral researcher at Technion. “The model integrates many subtle cues to generate a score that reflects both recurrence risk and expected benefit from chemotherapy.”
To create the tool, the researchers used deep learning methods to evaluate tumor regions and the surrounding tumor microenvironment to identify patterns linked to cancer behavior such as cell division, immune response, and tissue structure. This approach allows the model to estimate the Oncotype DX 21-gene recurrence score without requiring genomic testing.
“Genomic assays such as Oncotype DX have transformed adjuvant treatment selection for hormone receptor-positive, HER2-negative, early breast cancer but remain inaccessible to many patients because of high cost and logistical barriers,” the researchers wrote. The newly developed AI model addresses these barriers by using data already available in routine clinical workflows, offering faster times to prognosis and potentially making it available to all women with HER2-negative breast cancer.
The model was trained using a multimodal deep-learning framework that incorporated digital whole-slide images and clinical features. It leveraged a foundation model pre-trained on 171,189 histopathology slides, enabling it to learn generalizable visual features through self-supervised learning. The system was then fine-tuned using data from 8,000 patients enrolled in the TAILORx randomized clinical trial. This training approach was notable for its scale and for incorporating randomized trial data, which allowed researchers to evaluate not only prognostic performance but also predictive value for chemotherapy benefit.
This research builds on prior studies which have shown that features visible in histopathology images can correlate with genomic risk. Earlier studies demonstrated the potential of inferring cancer recurrence scores from H&E images, but these efforts were limited by smaller datasets and lacked validation using randomized clinical trial data.
To validate their model, the Technion team analyzed data from the TAILORx trial, one of the largest randomized breast cancer studies, and conducted external validation across six independent cohorts comprising more than 5,000 patients. The model demonstrated strong predictive performance, with an area under the curve of 0.898 for identifying high genomic-risk disease. It also showed the ability to stratify patients by recurrence risk and chemotherapy benefit, including stratification of subgroups based on whether chemotherapy should be avoided or would be of clinical benefit.
“These findings show that AI applied to routine histopathology can serve as a practical and scalable tool for guiding chemotherapy decisions in hormone receptor-positive, HER2-negative, early breast cancer,” the researchers wrote.
Use of the tool revealed some important distinctions between patients and their projected outcomes. For example, chemotherapy benefit was demonstrated in premenopausal patients classified as high risk by the model but not in postmenopausal patients classified as low risk. The model also reclassified some of patients who are considered high risk by current criteria into lower-risk categories, an important change for this subset of patients since chemotherapy can have a range of adverse side-effects.
The researchers noted that their work breaks new ground. “To our knowledge, this is the first digital pathology AI model for recurrence-score estimation to be assessed retrospectively using data from a [randomized clinical trial],” the researchers wrote. Taking this approach allowed for direct assessment of treatment benefit, which prior observational studies had not done.
Future steps include prospective validation studies to confirm real-world effectiveness while the team continues to train and refine the tool’s performance. The researchers are also considering expanding its use to other forms of cancer.
The hope is that the tool could eventually providing treatment insight to a broader range of patients, in particular those who may not have access to genomic testing. This tool only requires digitization of pathology slides. Using such images, prognosis and treatment advice can be generated in minutes. Patients with low scores from the tool could avoid chemotherapy, while those showing higher scores could be considered for treatment.
“This framework offers a pragmatic pathway for integrating AI-based risk assessment into standard decision making,” the researchers wrote.
