Graph Representation Learning and Explainability in Breast Cancer Pathology: Bridging the Gap between AI and Pathology Practice

According to the global cancer statistics released on December 14th, 2020, by the International Agency for Research on Cancer (IARC), the estimated new breast cancer cases in 2020 reached 2.26 million (24.5% of the total number of cancer cases for females), making female breast cancer the most commonly diagnosed cancer globally1, for the first time. While it is the leading cause for death in women with cancer, the survival rates for breast cancer are eminently high, subject to early diagnosis and adequate treatment. A study by [1] shows that for patients diagnosed with breast cancer during 2005–2009, the five-year survival rate rose to 85% or higher in 17 countries worldwide. This has been achievable due to early diagnosis through mass screening programs or intensive early diagnostic activity. Therefore, accurate, early diagnosis and risk assessment of breast cancer subtype for a patient holds a paramount role in the treatment and survival of the patient [2, 3]. In a clinical setting, a pathologist manually inspects a tissue specimen to detect and assess breast lesions, and thereby, to estimate any cancer intrinsic subtypes of the lesions given a predefined grading system. The subtypes confer different levels of risk according to their probability of transitioning to invasive carcinoma. For instance, lesions with atypia or ductal carcinoma in situ (DCIS) are associated with higher risks compared to benign lesions [4, 5]. The manual inspection by a pathologist starts with the discernment of coarse morphological and topological distribution attributes of the tissue, followed by the localization and analysis of specific regions of interest (RoIs). Further assessment is confined to the RoIs to analyze the phenotype and organizational properties of cells for subtyping the tissue specimen. Although diagnostic criteria for cancer subtypes are established, the continuum of histologic features phenotyped across the diagnostic spectrum prevents having clear decision boundaries between the cancer subtypes. Thus, manual inspection is a time-consuming process with significant intra-and inter-observer variability [5–7]. For instance, a study by [5] shows that the inter-pathologist agreement can be as low as 48% for breast lesions with atypia. The aforementioned challenges in manual diagnosis and the increasing incidence rate of breast