C-Strik: AI powered precision medicine tool for breast cancer

Grade is a measure of aggressiveness in breast cancer and is one of the main criteria used to make treatment decisions in breast cancer (BC), especially in young women. It can also predict survival and other outcomes for BC patients. Figures 2 and 3 show the current grading categories for BC using the Nottingham grading system that has remained unchanged over the past 50 years. There is significant inter and intra-observer variability in grading BC. Some large studies report as low as 30% concordance amongst pathologists in assigning specific grades to BC. This discrepancy in grading is because of lack of quantitative criteria and subjectivity involved in looking at the various parameters as shown in figures 3 and 4. For example, the variation in size and shape of the cancer nuclei is scored as mild, moderate or marked which can be very subjective.

My hypothesis is that adding objective criteria using machine learning can increase the grading accuracy in BC and be more reliable in guiding treatment decisions. Furthermore, it can help reduce the use of expensive multigene assays such as OncotypeDx in BC that cost upwards of 6000 CAD per test.

Adding a low cost alternative using the existing routine slides, rather than expensive tests to guide therapy can help women in parts of the world with less access and forbidding costs to get the same level of treatment as anywhere else. It can equalize care for everyone regardless of socio-economic status, geography and means to afford care.

Background reserach

I started by researching the studies in Pubmed that were available on AI assisted grading in breast cancer. Majority of the papers were focused on diagnosing breast cancer, and only a handful of papers focused on using AI for grading. These studies trained their model on only a small number of slides and were not in clinical use anywhere in the world.

Next, I started assesing how to create an AI model to assess the BC grade. I found out that there are large Foundational models eg. CONCH, Virchows and Gigapath that were trained on pathology datasets to identify hundreds of parameters (link below). These could be further customized to help with my question. I went through the pros and cons of each of these models and decided upon Virchows for my specific use (Fig 1). I deployed the model on Google Colab. Next I downloaded 30 annonymized BC slides from Queen's University database for my initial training set (Fig 2). These images were annotated for different components and also had grade labels as shown in fig 1. Each slide was broken into 224 x 224 pixel patches for analysis. My model goes through 3 iterations: #1 - tubule grade; #2 - nuclear grade; and #3 - mitotic rate fig --. The layers are put together through multi-instance learning (MIL) to give a final grading for each patch and then slide-level grading.

I obtained the open source BC pathology images from The Cancer Genome Atlas (TCGA) image archives. This dataset has genetic and outcome information associated with the cases for my future work. I applied my Virchows model on these 458 TCGA slides to test the accuracy.

https://github.com/georg-wolflein/pathology-foundation-models

https://www.cancerimagingarchive.net/collection/tcga-brca/

Results

My AI model, C-Strik was dveloped on a pathology foundation model, Virchow that uses 1.5 million pathology images. I trained specifically for breast cancer grading using training set of 30 cases from Queen's University and then tested it on almost 500 TCGA breast cancer images. These slides generated 2,240,000 patches. Each patch was assigned a grade for the 3 components and then an overall grade score out of 9.

C-strik provides explainable AI solution for quantitative and objective grading of breast cancer with overall 76% accuracy. My model is better than most studies with this type of very complex models, because of multi-layered neural networks. C-strik is trained to be an associate to the pathologist to increase the grading accuracy at the moment as per my hypothesis that AI + human is better than AI alone or human alone.

My results show an AI model can be trained to effectively grade breast cancer. It outputs objective criteria that the model uses to assign a score to individual components, eg. tubule density, coarseness of the nuclear membrane and number of mitoses per high power fields. These are explainable features that the pathologists can understand and assess under a microscope. The explainable quality of the model will help with the adoption when I deploy it in the clinical setting.

I want to add more images from diverse patient populations by partnering with Queen's and KHSC, and other hospitals. It is known that AI starts drifting on different populations and so I want to make it robust by adding many different datasets.

Adding more images will increase the accuracy of my model, so that eventually it can be fully autonomous.

I want to correlate the breast cancer grade scores from my model with genetic test scores such as Oncotype Dx to assess if the AI grading can offer a reliable alternative for treatment decisions instead of the costly tests that are not available in many parts of the world.

By doing this, I want to equalize breast cancer care for every woman in every part of the world.

I am doing an internship at the Center for Advanced Computing this summer and that experience will help me futher refine my model.

I have joined the Y-combinator online start up school to commercialize my invention

I will again work with Queen's computing school as I did in the past summer to gain more in-depth knowledge of AI models.

I would like to thank my amazing mentors, Dr. Gabor Fichtinger and Rebecca Hissey; and the entire Perk lab for an invaluable summer opportunity to learn AI tools. A special thanks to Josh for being my lab coach!

My future mentors at the CAC, I am looking forward to an amazing internship program. Thank you for giving me the opportunity.

My mom, Dr. Sonal Varma for always helping me and putting up with me! And my Dad for believing and assuring my brother and I, that age should never stop us from doing what we want to do. Finally, to my little brother for all sorts of candid comments and questions about the project that I am fully prepared for anything that comes my way :-)

Vorontsov, E., Bozkurt, A., Casson, A., et al. (2024). A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine, 30, 2924–2935.

Chaudhary, N., & Dhunny, A. Z. (2025). An artificial intelligence model for early-stage breast cancer classification from histopathological biopsy images. Frontiers in Artificial Intelligence, 8, 1627876.

Kaczmarzyk, J. R., Van Alsten, S. C., Cozzo, A. J., et al. (2026). Towards interpretable prediction of recurrence risk in breast cancer using pathology foundation models. npj Digital Medicine, 9, Article 149.

Dalton, L. W., Pinder, S. E., Elston, C. E., Ellis, I. O., Page, D. L., Dupont, W. D., & Blamey, R. W. (2000). Histologic grading of breast cancer: Linkage of patient outcome with level of pathologist agreement. Modern Pathology, 13(7), 730–735.

Elston, C. W., & Ellis, I. O. (1991). Pathological prognostic factors in breast cancer: The value of histological grade in breast cancer: Experience from a large study with long-term follow-up. Histopathology, 19, 403–410.