Artificial Intelligence (AI) - Based Models for Biomarkers Prediction

Despite the improvements in therapeutic approaches, breast cancer remains the most commonly occurring cancer and leading cause of cancer death in women worldwide. Breast cancer treatment decisions are based mainly on the presence or absence of established prognostic and predictive biomarkers, mainly the hormone receptors, estrogen (ER) and progesterone (PgR), the human epidermal growth factor receptor 2 oncoprotein (HER2) in correlation with ki67 proliferation index. Currently, PCR-based assays, next-generation sequencing (NGS), and immunohistochemistry (IHC), followed by in situ hybridization (ISH) when needed, are used for biomarker testing. Ιn an effort to overpass technical limitations, human subjectivity which may yield different results between different observers leading to inappropriate treatment and complications that can adversely affect patient quality of life, alternative, cost-effective approaches using AI-based digital pathology analysis of histopathology images has received significant attention last years. Digital pathology uses whole-slide image scanners and combined with artificial intelligence (AI) algorithms intend to improve diagnostic accuracy, enabling rapid patient stratification and optimal treatment regimens, thereby improving patient outcomes. Identification of invasive breast tumors and lymph node metastasis, evaluation of hormonal status, breast cancer grading and mitotic count evaluation, biomarkers identification based on images of the tumor microenvironment have also been enhanced by AI quantitative analysis. AI has promising results so far. More studies are, however, needed in order to overcome obstacles and before it is validated for patient care in the clinical setting. DL technology has been implemented in clinical pathology in an attempt to automate the way pathologists evaluate hematoxylin and eosin (H/E) stained slides, but also to attain information such as molecular background and protein expression, given that genetic variations could likely be reflected in the morphology of cancer cells, and although not detectable to the human eye, these changes can be recognized by advanced DL algorithms and then correlated with a specific molecular variation or protein expression. For the purpose of our study, we will use FFPE tumor blocks from patients enrolled in 7 randomized and observational studies conducted by the Hellenic Cooperative Oncology Group (HeCOG) to define via NGS their underlying mutational profile. Subsequently, a DL algorithm will be trained in order to examine if any genetic alterations found in breast cancer could be detected by the DL model in H/E whole section slides, which would suggest that these variants reflect in the morphology of the cancer cells. Accordingly, we will explore if our algorithm could be trained to identify tumor infiltrating lymphocytes (TILs) by a similar approach using the annotated by our pathologist TILs regions on the H/E slides, as despite their importance, TILs scoring seems to be differentiated among pathologists because of substantial interobserver variation. This fact raises an urgent need for the exploration of novel methods for a more accurate scoring system. Indeed, the International Immuno-Oncology Working Group suggested a computational assessment of TILs based on deep learning models. The evaluation of biomarkers related to TILs, such as CD8, in IHC-stained images has also shown significant potential when DL approaches are used, yielding promising results in no biased TILs evaluation and providing us with valuable details about their distribution and spatial relationships . Since there are still limitations and things to enhance and more data and parameters need to be incorporated to create more accurate and powerful models, we aim to improve TILs evaluation and therapeutic response prediction in breast cancer, through our deep learning model with the use of H/E whole slide images and digitalized IHC-stained CD8 images for lymphocytes' automatic detection for better quantification of immune response. Additionally, and apart from ER and PgR IHC-stained slides digitalization, we intend also to digitalize our archive's breast cancer IHC sections stained with HER2 ab, which have already been evaluated by our experienced pathologist, as well as HER2/TOP2A/CEN17 FISH slides. Subsequently, we intend to develop an automated deep learning (DL) - based image analysis algorithm with the ultimate goal of HER2 evaluation with a more accurate than human scoring manner, with the advantage of scoring every cell within a sample specifically to better identify patients with low-level HER2 expression given that T-DXd seems to be efficacious in this patient population . In this way, we will improve diagnostic accuracy by enabling pathologists to review high-resolution digital slides, apply advanced image analysis algorithms, and reduce inter-observer variability, for an accurate HER2 expression in breast cancers.

Sex

Female

Minimum age

18 Years

Inclusion Criteria:

* Age 18 and above
* Histologically confirmed BC
* All treated with adjuvant dose-dense sequential chemotherapy (dds-CT)
* Tumor tissue specimen (FFPE) availability

Exclusion Criteria:

* not adequate, and unsuitable tissue for IHC, FISH and analysis

• BACKGROUNDVandenberghe ME, Scott ML, Scorer PW, Soderberg M, Balcerzak D, Barker C. Relevance of deep learning to facilitate the diagnosis of HER2 status in breast cancer. Sci Rep. 2017 Apr 5;7:45938. doi: 10.1038/srep45938. PMID 28378829
• BACKGROUNDSoliman A, Li Z, Parwani AV. Artificial intelligence's impact on breast cancer pathology: a literature review. Diagn Pathol. 2024 Feb 22;19(1):38. doi: 10.1186/s13000-024-01453-w. PMID 38388367
• BACKGROUNDSaldanha OL, Loeffler CML, Niehues JM, van Treeck M, Seraphin TP, Hewitt KJ, Cifci D, Veldhuizen GP, Ramesh S, Pearson AT, Kather JN. Self-supervised attention-based deep learning for pan-cancer mutation prediction from histopathology. NPJ Precis Oncol. 2023 Mar 28;7(1):35. doi: 10.1038/s41698-023-00365-0. PMID 36977919
• BACKGROUNDPrat A, Parker JS, Fan C, Perou CM. PAM50 assay and the three-gene model for identifying the major and clinically relevant molecular subtypes of breast cancer. Breast Cancer Res Treat. 2012 Aug;135(1):301-6. doi: 10.1007/s10549-012-2143-0. Epub 2012 Jul 3. PMID 22752290
• BACKGROUNDFiorin A, Lopez Pablo C, Lejeune M, Hamza Siraj A, Della Mea V. Enhancing AI Research for Breast Cancer: A Comprehensive Review of Tumor-Infiltrating Lymphocyte Datasets. J Imaging Inform Med. 2024 Dec;37(6):2996-3008. doi: 10.1007/s10278-024-01043-8. Epub 2024 May 28. PMID 38806950
• BACKGROUNDEl Nahhas OSM, van Treeck M, Wolflein G, Unger M, Ligero M, Lenz T, Wagner SJ, Hewitt KJ, Khader F, Foersch S, Truhn D, Kather JN. From whole-slide image to biomarker prediction: end-to-end weakly supervised deep learning in computational pathology. Nat Protoc. 2025 Jan;20(1):293-316. doi: 10.1038/s41596-024-01047-2. Epub 2024 Sep 16. PMID 39285224
• BACKGROUNDEl Nahhas OSM, Loeffler CML, Carrero ZI, van Treeck M, Kolbinger FR, Hewitt KJ, Muti HS, Graziani M, Zeng Q, Calderaro J, Ortiz-Bruchle N, Yuan T, Hoffmeister M, Brenner H, Brobeil A, Reis-Filho JS, Kather JN. Author Correction: Regression-based Deep-Learning predicts molecular biomarkers from pathology slides. Nat Commun. 2024 Feb 29;15(1):1868. doi: 10.1038/s41467-024-46298-5. No abstract available.