AI Predicts Who Benefits from Chemotherapy in Early Breast Cancer Using Routine Pathology Slides

A multicentre study found that an AI model using routine H&E pathology slides and clinicopathological data could estimate genomic recurrence risk and predict chemotherapy benefit in hormone receptor-positive, HER2-negative early breast cancer. In the TAILORx test set, the model achieved an AUC of 0.898 for identifying high-risk disease and remained robust across 5,497 external cases, supporting a fast, scalable alternative to genomic assays.

In early-stage, hormone receptor-positive, HER2-negative breast cancer, one of the hardest decisions comes after surgery: who actually needs chemotherapy. The stakes are high on both sides. Chemotherapy can reduce recurrence risk, but many patients derive little or no benefit and still face neuropathy, fatigue,  and other lasting toxicities. Genomic assays such as Oncotype DX have helped refine that decision, but their cost, turnaround time, and uneven global availability remain major barriers.

The study, published in The Lancet Oncology, suggests that artificial intelligence may offer a more accessible approach. Researchers from the Technion, working with collaborators in the US, Europe, and Australia, report that this model could provide a practical alternative. The findings show that deep learning applied to routine H&E-stained pathology slides, combined with clinicopathological variables, can estimate genomic recurrence risk and identify subgroups more or less likely to benefit from chemotherapy.

How the Model Works

The system works on material that is already generated during standard diagnosis. After a tissue sample is scanned, the AI analyzes multiple regions of the tumor and its microenvironment, searching for patterns linked to proliferation, immune response, tissue architecture, and treatment sensitivity.

“These are complex biological signals that the human eye cannot consistently quantify,” said Dr. Gil Shamai of the Technion’s Geometric Image Processing Laboratory, who led the study. “The model integrates many subtle cues to generate a score that reflects both recurrence risk and expected benefit from chemotherapy.”

The AI model was developed using digital whole-slide images and clinical data from patients with hormone receptor-positive, HER2-negative invasive breast cancer. It was built on a foundation model pre-trained on 171,189 histopathology slides, then fine-tuned and validated using samples from the landmark TAILORx randomized trial. After quality control, the investigators analyzed 8,284 patients, with performance assessed in a TAILORx test set of 2,407 patients and then externally validated in six independent cohorts totaling 5,497 patients.

In the TAILORx test set, the AI model classified 45.6% of patients as low risk, 42.4% as intermediate risk, and 12.0% as high risk. For identifying high genomic-risk disease, defined as a recurrence score of 26 or higher, the model achieved an AUC of 0.898. External validation showed similarly strong performance, with AUC values ranging from 0.858 to 0.903 across diverse institutions and scanner settings. That consistency matters, because digital pathology tools often lose accuracy when moved beyond the environment in which they were trained.

More importantly, the model did more than mirror a genomic test. By leveraging the randomized design of TAILORx, the team also examined whether the AI output could predict actual chemotherapy benefit. That is a far more clinically meaningful endpoint than recurrence risk alone.

“Using data from a randomized trial allowed us to test whether the model truly predicts benefit from chemotherapy, not just recurrence risk,” said Dr. Shamai.

The findings were especially relevant in subgroups where treatment decisions are often difficult. Chemotherapy benefit was evident in premenopausal patients classified by AI as high risk, with a hazard ratio of 0.63. By contrast, postmenopausal patients classified by AI as low risk showed no meaningful chemotherapy benefit, with a hazard ratio of 0.94. Among clinically high-risk postmenopausal women defined by MINDACT criteria, 31.3% were reclassified by the model as low AI risk, a group that also showed no benefit from chemotherapy. The authors argue that this kind of refinement could help spare overtreatment in routine practice.

Prof. Ron Kimmel, head of the laboratory in the Henry and Marilyn Taub Faculty of Computer Science, described the logic more simply:  “Instead of testing genes, we look directly at the tissue. Just as eye color can be determined by looking at the eyes rather than analyzing DNA, our system extracts a visual signature from pathology images that informs optimal treatment.”

“This is the first AI model shown to predict treatment benefit in breast cancer directly from pathology samples,” said Prof. Dvir Aran of the Technion’s Faculty of Biology, a co-leader of the study.

According to the study, the model may be especially useful where genomic infrastructure is limited. It requires no additional tissue, no specialized molecular workflow, and no long wait for results. 

“In developing countries, where genomic testing is largely unavailable, this tool could dramatically expand access to personalized cancer care,” said Prof. Aran. “In high-income countries, it could reduce costs, shorten diagnosis time, and improve decision accuracy.”

The researchers stress that the model is not yet a wholesale replacement for standard care in every setting. For some node-negative patients outside the clearest high- and low-risk groups, conventional genomic testing and clinicopathological assessment still remain important. But the broader message is clear: routine pathology, long viewed mainly as a diagnostic snapshot, may also contain enough information to guide treatment intensity.