Medizinische Universität Graz Austria/Österreich - Forschungsportal - Medical University of Graz
Gewählte Publikation:
Boemcke, C.
Probability Prediction of Homologous Recombination Deficiency in Serous Ovarian Cancer Through a Neural Network Based on Histopathological Slides from TCGA
Humanmedizin; [ Diplomarbeit ] Medical University of Graz; 2021. pp. 56
[OPEN ACCESS]
FullText
- Autor*innen der Med Uni Graz:
- Betreuer*innen:
- Holzinger Andreas
- Regitnig Peter
- Altmetrics:
- Abstract:
- Introduction: Ovarian cancer is one of the most frequent causes of death worldwide. Due to a long asymptomatic course of the disease, the diagnosis is often only made in very advanced stages and the prognosis is mostly unfavourable. Serous ovarian carcinoma, which can occur as low-grade and high-grade, is the most common morphological type. One of the causes for the development of this malignant tumour is a disturbance in one of the repair mechanisms of the DNA, namely homologous recombination. If this repair mechanism of the cell fails, inefficient DNA repair occurs, which promotes the development of cancer. Patients with BRCA1 or BRCA2 mutations are genetically deficient in homologous recombination and therefore often suffer from breast or ovarian cancer. Homologous recombination deficiency can also be caused by additional factors, such as other genetic mutations, DNA methylation or factors that are still currently unknown. This deficiency and specifically intervene with so-called PARP inhibitors therapeutically can be used. Those prevent DNA repair by inhibiting the base excision repair pathway (BER) and force the cell to resort to homologous recombination as a second repair mechanism. However, if this also does not work, as in case of BRCA mutants, the DNA of the cell perishes and hence tumour growth is inhibited. The prerequisite for the effectiveness of PARP inhibitors is therefore a homologous recombination deficiency (HRD). Histologically, this deficiency may be seen in the different morphology of the cells. Up to date, however, only limited research has been undertaken on a potential correlation of the visual characteristics of cells and its HRD status. In order to address this, artificial intelligence provides a powerful tool to automatically analyse and classify image data. This work investigates whether it is possible to automatically correlate and predict HRD status based on microscopic images by means of neuronal networks. Methods: Microscopic whole slide images (WSI) of 436 patients with ovarian cancer were analysed, together with mutation data from the GDC portal TCGA. HRD status was defined based on a mutation in BRCA1, BRCA2, CHEK1 or PTEN. A neural network was trained to predict the presence of HRD using these images. Results: With an accuracy of only 45.3%, the neural network was not able to predict whether an HRD is present or not. However, the classification between tumour and normal tissue succeeded with a hit accuracy of 83.9%. Conclusion: In this work, it was shown in principle that it is possible to use artificial intelligence to find morphological correlates in histological specimens. The classification of HRD in this project was not successful.