Image analysis of immunohistochemistry-based biomarkers in breast cancer
von Judith Burchhardt
Datum der mündl. Prüfung:2022-11-21
Erschienen:2022-11-18
Betreuer:PD Dr. Peter Middel
Gutachter:PD Dr. Peter Middel
Gutachter:Prof. Dr. Günter Emons
Gutachter:PD Dr. Sabine Sennhenn-Kirchner
Dateien
Name:Dissertation_ohne_CV.pdf
Size:8.08Mb
Format:PDF
Zusammenfassung
Englisch
Breast cancer is the leading form of cancer in women in Germany with about 69.000 new cases anually and a lifetime risk of 12.9%. One in eight women will develop a malignant neoplasm of the breast. Early diagnosis measures include regular clinical examinations, mammography and biopsies of suspicious lesions. Definite diagnosis is based on histopathology. Pathologic work-up encompasses histomorphology on H&E stained slides and a set of four biomarkers that provide prognostic information about the expected course of disease and predictive information about the likeliness to benefit from different clinical treatments. In breast cancer, immunohistochemistry is currently the most common type of biomarker. Immunohistochemistry conventionally relies on manual histological interpretation, but automated techniques based on image analysis have become increasingly available. In the present study, n = 613 breast cancer core needle biopsies from a single pathological laboratory (Pathologie Nordhessen, Kassel, Germany) were re-analysed by whole slide scanning of the histological specimens and image analysis of the biomarkers estrogen receptor, progesterone receptor, Her2 receptor and Ki-67 by the software package QuantCenter (3D Histech). The results were compared to manual biomarker interpretation by board-certified pathologists. Digitisation of the histological slides by a state-of-the-art tile scanner (3D Histech Pannoramic P250 Flash II) required 82 seconds per slide on average (standard deviation: ± 38s) and seemed technically mature. Allocation and storage of the large files constitute major issues that require costumised solutions. Image analysis did not work with out-of-the-box settings but required optimisation on local cases. After training of the software, satisfying rates of concordance were achived for estrogen and progesterone receptors with Cohen's kappa coefficients of κ = 0.86 and κ = 1.0. In Ki-67, systematic differences between manual scoring and image analysis were noticed and the best concordance achieved was κ = 0.68. Her2 yielded a good concordance of κ = 0.74 in a training set of n = 19 representative cases but only a moderate concordance of κ = 0.55 in the complete cohort. Exploratory analysis of Her2 yielded additional information on the physical basis of manual Her2 scoring. The findings indicate that image analysis is a mature technique that can be used to supplement the analysis of biomarkers in breast cancer. Image analysis has potential to decrease interobserver variance and to allow more precise quantitation. Yet, current software approaches require specific optimisation on local cases. The achieved concordance results from the representativeness of these training cases, which raises the question of how to define such reference standards. A possible solution could be centrally defined testing materials, for example tissue cultures with fixed levels of biomarker expression, that could be used for standardised local optimisation.
Keywords: image analysis; breast cancer; immunohistochemistry (IHC); biomarkers; magnification rule; digital pathology (DP); digital image analysis (DIA)