Deep learning-based computational pathology for breast cancer precision medicine

Background

Cancer is a leading cause of death globally. To improve care and patient outcomes for cancer patients, there is a need for fast, reliable and cost-effective precision diagnostic solutions accessible for all patients.  Manual histopathology assessment of tumour tissue remains the gold standard to detect, diagnose and characterise cancer tumours in clinical routine. However, there is an increasing need to stratify patients more precisely to ensure that toxic treatments are used only in those patients who stand to benefit. Molecular profiling, currently the dominating approach for patient stratification in precision medicine, is highly expensive which limits broad patient access and imposes a high economic burden on healthcare systems. Advanced AI-based image analysis, primarily by means of deep learning, of histopathology whole slide images (WSI) offers novel means for precision diagnostics and cancer patient stratification that are information-rich, precise and rapid, while remaining cost-effective[1]. Central to their validation is access to large datasets with accurate patient level data and long term outcomes.  The University of Edinburgh group host an extensive tissue biobank from randomised phase III clinical trials with well-annotated outcome data held by the Early Breast Cancer Trialists’ Collaborative Group (EBCTCG) global consortium. In addition, within Scotland’s renowned, high quality routine electronic health records, the Scottish Cancer Registry provides an excellent resource to investigate the real world cohort of NHS Lothian patients who are cared for by the Edinburgh Cancer Centre [2]. Including a real world cohort in model validation allows and understanding of how AI-based models perform in the face of competing risks such as comorbidity that is under-represented in clinical trials.

Aims

Working in partnership with the EBCTCG and the Karolinska Institute (KI), we propose a structured approach to the validation of the next-generation AI-based tools using patient level meta-analysis and conventional pathology information. The project is divided into an initial data-generation section, followed by the computational analysis and interpretation of the data.  A key aim is to validate models in both an EBCTCG clinical trial dataset and an Edinburgh Real World dataset.

1. Validation of the prognostic performance of an existing AI-based CE-IVD marked solution for breast cancer risk stratification (Stratipath Breast) that was prototyped at KI [3]. Validation will be performed in large real-world data in the form of a population-based large retrospective cohort in UK/Scotland.    

2. Development of a nomogram for survival prediction in primary breast cancer patients that combines established clinicopathological factors with AI-based risk stratification.     

3. Development and validation of a deep learning -based model (e.g deep convolutional neural network) for Estrogen Receptor (ER) status prediction directly from H&E stained histopathology whole slide images using stain-guided learning.

Training outcomes

This PhD project will provide broad research training in the use of AI and machine learning for cancer precision medicine. The project is designed to include both methodological components (deep learning -based medical image analysis; computational pathology), as well as components relating to clinical translation and establishing evidence in real-world data.

The student will develop broad knowledge and systematic understanding of the following areas: Cancer precision medicine, artificial intelligence and machine learning, statistical methodology for analysis of epidemiological studies, histopathology. A link with the Edinburgh Breast Unit will allow contextual understanding of the clinical problem.

Deep knowledge and expertise is expected to be developed in the areas of: deep learning and its application in histopathology image analysis, development and validation of prediction models designed for patient stratification in clinical applications, breast cancer histopathology.

References

1.       Acs, B., M. Rantalainen, and J. Hartman, Artificial intelligence as the next step towards precision pathology. Journal of Internal Medicine, 2020. 288(1): p. 62-81.

2.         https://cancer-data.ecrc.ed.ac.uk/.

3.       Wang, Y., et al., Improved breast cancer histological grading using deep learning. Ann Oncol, 2022. 33(1): p. 89-98.