From Research to Clinical Diagnostics: Developing and Validating Biomarkers and Artificial Intelligence for Pathology

Abstract

In 2021, digital pathology was deployed in the hospitals of the western health region of Norway. Histological tissue specimens previously viewed under the microscope on glass slides, are now being scanned, and whole slide images (WSIs) are viewed digitally. Digitisation enables the use of advanced technologies to take over repetitive and timeconsuming tasks such as biomarker quantification. Furthermore, digital image analysis (DIA) and artificial intelligence (AI) can be used to perform complex tasks such as pattern recognition and classification, to assist healthcare professionals. The research presented in this thesis aims to explore methods which may improve current diagnostic and prognostic guidelines for breast cancer and endometrial hyperplasia in pathology. To challenge current limitations of visual assessments and investigate if addition of quantitative methodology and AI-assistance tools can improve reproducibility and accuracy of diagnosis and prognosis. The end goal to reduce the risk of under- and over-treatment of these patients. In Norway, 3,000 to 4,000 women will be diagnosed with endometrial hyperplasia every year. This condition is characterised by the excessive proliferation of endometrial glands in the uterine lining. The diagnosis of endometrial hyperplasia has undergone several important evolutions in recent decades. However, the prognostic evaluation, to assess the likelihood of this condition to progress to endometrial cancer, is still limited by subjective visual assessment of tissue morphology. In the first study, the biomarkers PTEN and PAX2 were evaluated for their prognostic value in endometrial carcinogenesis. A quantitative method assessing PAX2 protein expression revealed prognostic separation of patients diagnosed with endometrial intraepithelial neoplasia with low- and high-risk of progression to cancer. In a second study, an AI-based tool was developed, to detect and quantify morphological features of endometrial hyperplasia. The tool (ENDOAPP) was able to identify patients with low-risk and high-risk for progression. Furthermore, its accuracy was equal to and marginally superior to a semi-quantitative morphometric method (D-score) and traditional visual classifications (WHO94, WHO20, EIN), respectively. To state that the diagnosis and treatment of cancer has a long history would be an understatement. The arrival of new technology, molecular advances and AI continues to revitalise the way cancer is viewed in the clinic. The measurement of proliferation in breast cancer has undisputed prognostic implications. However, quantification of proliferation markers is controversial citing lack of standardisation. AI may provide a promising solution for the establishment of improved methods for objective, automated, reproducible quantification of proliferation markers such as mitotic count and Ki67. In the third study, in-house and commercial DIA tools were investigated alongside manual Ki67 quantification methods for their prognostic capability and variability. It was observed that DIA tools were superior to their manual counterparts with regards to their discriminative ability for separation of low-risk and high-risk for distant metastasis free progression. Furthermore, the cut-offs currently used for binary risk categorisation of proliferation markers should be carefully re-evaluated if we wish to standardise quantification of Ki67. In the final study, a deep learning tool was investigated for the detection and quantification of mitotic count in several cancers, including breast cancer. It was observed that automated mitotic count was prognostic in multiple cancer types in addition to breast cancer, where it is routinely performed. It is important to emphasise that the results presented in this thesis are limited to the datasets presented. These were retrospective datasets, and often confined to a single hospital, with the exception of the fourth study. Therefore, the methods run the risk of overfitting and hidden bias. It is therefore imperative that these tools are validated in external datasets to ensure their robustness, uncover any bias or overfitting, and to confirm their prognostic validity. Although the studies presented in this thesis suggest the validity of investigating AI-tools for clinical use, further study to critically evaluate their worth is still required.