Abstract
Digital pathology (DP) and whole slide images (WSIs) are rapidly becoming the first option for routine diagnostics. Successful application of artificial intelligence (AI) methods such as machine learning algorithms (ML) to WSIs have the potential to create new supportive diagnostic clinical tools that can improve diagnostic accuracy, reproducibility and objectivity and provide new insights into human and canine cancer.
This thesis focused on the detection and characterisation of canine soft tissue sarcomas (cSTSs) using ML. Currently, the diagnosis of cSTSs is based on histological assessment. In particular, by assessing certain histological features such as the degree of differentiation, necrosis score and mitotic score, it is possible to define a final tumour grade, which aids in prognostication for patients. Due to the subjectivity of the scoring system, grade disagreements are reported in human and cSTS cases. These results were confirmed in our study where we found only a fair level (κ = 0.36) of diagnostic concordance between pathologists in grading these tumours illustrating the need for automated image analysis tools. In order to achieve this goal, the first essential step was to create an appropriate and comprehensive digital slide dataset. The study gathered a large-scale dataset of 1166 histopathological WSIs of cSTSs (n=752), canine mast cell tumours (MCTs) (n=359) and canine apocrine gland anal sac adenocarcinoma (AGASACAs) (n=55) with related clinical information. Once the slide panel was assembled, the study focused on providing a tool for automatic detection of tumour necrosis and mitosis in annotated histopathological WSI of STS using ML. The ML algorithm applied (DenseNet161) had an accuracy for necrosis and mitosis detection of 92.7% and 50.0%, respectively. In conclusion, the results presented here have demonstrated that digital pathology and ML algorithms can potentially be used as a diagnostic support tool for the detection and characterisation of cSTSs.