Triple negative breast cancer (TNBC) is an aggressive subtype accounting for 10-15% of all breast cancers. It lacks estrogen receptor, progesterone receptor, and HER2 expression, making it particularly difficult to treat and associated with a high risk of recurrence after treatment.
Neoadjuvant chemotherapy (NAC) is systemic treatment given before surgery with the goal of shrinking the tumor, converting inoperable disease to operable, and in some cases allowing less extensive surgery than mastectomy. About 40-50% of TNBC patients treated with NAC achieve a pathological complete response (pCR), meaning no remaining tumor is found at surgery.
A key problem is that patients who do not respond well still endure roughly six months of toxic chemotherapy with no benefit. Being able to predict in advance who will and will not respond could spare non-responders the side effects and allow them to proceed directly to surgery, improving both outcomes and quality of life.
Existing biomarkers used to predict NAC response in TNBC include tumor size, histological grade, Ki-67 proliferation score, and tumor-infiltrating lymphocytes (TILs). Genetic tests like Oncotype DX and MammaPrint also exist but are expensive, time-consuming, and not applicable to TNBC.
This study aimed to use deep learning technology to predict NAC response from the visual patterns in standard pre-operative tumor biopsy slides, stained with hematoxylin and eosin (H&E). H&E staining is the routine method used by pathologists to examine tissue under a microscope.
The approach uses whole slide images (WSIs) - digitally scanned microscope slides - fed into a convolutional neural network (CNN). The CNN was trained to identify subtle morphological patterns linked to treatment response, learning from the images without being told explicitly which features to look for.
This proof-of-concept study sought to determine whether the visual data encoded in a pre-treatment biopsy already contains enough information to reliably predict whether a patient will have a good, moderate, or poor response to chemotherapy - effectively asking whether the tumor's appearance foretells its behavior under treatment.
The study collected 221 H&E-stained biopsies from 205 patients diagnosed with TNBC of no special type (NST) from five hospitals between 2017 and 2022. Cases were required to have received at least 4.5-6 months of NAC before surgery. Slides from earlier years were re-stained to correct faded staining.
A pathologist manually drew precise boundaries around invasive tumor areas in each slide, including a small rim of surrounding tissue but excluding necrosis, extensive fibrosis, artifacts, and ductal carcinoma in situ (DCIS). This precise manual segmentation provided the training regions for the CNN.
Cases were divided into three response groups based on the EUSOMA scoring system applied to the surgical specimen: good response (less than 10% residual tumor), moderate response (10-50%), and bad response (more than 50% residual tumor). Because moderate and bad responders were relatively few, these groups were merged for training to improve discrimination.
The CNN was trained using image patches of 200x200 micrometers at 40x magnification on the Aiforia cloud-based platform. A 5-fold cross-validation was performed on cases from 2017-2021, and the final model was tested on a completely separate set of 52 biopsies from 50 patients from 2022.
The model's performance was measured using the area under the receiver operating characteristic curve (AUC ROC), a standard measure of how well a model distinguishes between classes. An AUC of 1.0 represents perfect discrimination, while 0.5 represents chance-level performance.
In cross-validation across the training set, the model achieved AUC values ranging from 0.541 to 0.730 with a mean of 0.616. On the independent test set of 52 new biopsies, the model achieved an AUC ROC of 0.696 with a 95% confidence interval of 0.532-0.861.
In practical clinical terms, applying this model could identify 23.1% of moderate and poor responders while falsely classifying only 2.6% of good responders as poor responders. This suggests meaningful clinical utility for identifying a subset of patients who would not benefit from NAC.
The achieved AUC of 0.696 outperforms a published benchmark of 0.63 based on structured clinical data including histological grade, TILs, and Ki-67 - demonstrating that the image-based AI approach captures predictive information beyond what standard biomarkers alone provide.
By analyzing regions where the AI made predictions with high confidence (above 90%), researchers identified visual patterns associated with each response category. These histological features provide biological insight into what the neural network learned to recognize.
Features suggestive of good response to chemotherapy included syncytial cell groups with minimal tubule formation, biphasic populations of epithelial cells, and a prominent presence of leukocytes (lymphocytes, plasma cells, neutrophils) sometimes with nuclear debris and apoptotic cells - collectively reflecting an active immune response within the tumor.
Features associated with poor response included relatively abundant sclerotic or desmoplastic stroma, spindle cell nuclei in low-cellularity areas, compact blue round epithelial zones, and areas with lower leukocyte density compared to surrounding tumor regions. These patterns suggest a stiff, immune-excluded tumor microenvironment.
These findings align well with established biological understanding: high TIL infiltration is a known favorable prognostic factor in TNBC, and dense fibrotic stroma has been associated with chemotherapy resistance. The AI appeared to confirm and extend known patterns rather than invent entirely novel ones.
The authors compared their results with several prior AI-based NAC prediction studies. Li et al. achieved an AUC of 0.847 but included multiple breast cancer subtypes and used automated tumor segmentation. Naylor et al. achieved mean AUCs of 0.654 and 0.601 in TNBC, comparable to the present study's 0.616 in cross-validation.
A key methodological difference in this study was training the CNN in a naive or agnostic manner - without directing it toward known biomarkers like TILs or Ki-67. This approach allows the model to potentially discover unknown predictive features in addition to confirming known ones, though it also means the model's reasoning is less transparent.
The authors made deliberate decisions to exclude DCIS, extensive necrosis, large fibrotic zones, and tissue artifacts from the tumor segmentation used for training. These exclusions were designed to prevent the CNN from learning patterns associated with secondary tumor changes rather than the primary biology that predicts treatment response.
A limitation acknowledged is that the exact chemotherapy regimen and number of cycles was not known for most patients. Treatment protocol knowledge could refine predictions, and a higher AUC might be achievable with more precise clinical data combined with image features. The study also lacks a detailed analysis of failure modes and misclassified cases.
The study envisions AI-driven tumor slide analysis as a future tool for therapeutic decision-making at the time of initial biopsy. A favorable AI prediction could support proceeding with NAC followed by surgery, while an unfavorable prediction might guide physicians to pursue surgery first followed by adjuvant chemotherapy instead.
The CNN analysis produces a result in seconds from a standard H&E slide, compared to days or weeks for genetic tests and significant pathologist time for manual grading. If validated at larger scale, this approach could be far more accessible and cost-effective than existing molecular diagnostic tests.
The authors suggest that future AI models incorporating both imaging and detailed clinical variables - including tumor size, lymph node status, and treatment regimen - could substantially improve prediction accuracy and potentially make traditional histological grading, TIL scoring, and expensive gene expression profiling partially redundant.
This proof-of-concept study demonstrates that deep learning analysis of standard H&E-stained pre-operative biopsies can extract predictive information about NAC response in TNBC, achieving an AUC of 0.696 that outperforms conventional clinical biomarker prediction.
The results validate the concept that microscopic tumor morphology encodes biologically meaningful information about how a tumor will behave under chemotherapy treatment - information that is accessible to AI even when not fully visible to trained human pathologists.
Future work should include larger datasets with more detailed clinical information, investigation of unknown biomarkers beyond those currently recognized, and rigorous analysis of model failures to move this approach from proof-of-concept toward clinical deployment.