Aleksandra Osowska-Kurczab
supervisor: Tomasz Markiewicz
Medical imaging is significant in tumour diagnosis and treatment planning, especially in abnormality localisation and early-stage examinations. Due to the unsatisfactory performance of pure radiology, additional histopathological tests are conducted to confirm the final diagnosis. Though this approach yields satisfactory results from the patient perspective, the whole process lacks cost and time optimisation, causing hampering and extensive expenditures on diagnostic procedures.
The universality of medical imaging and its wide popularity constitute considerable potential in providing adequate and direct treatment, especially in expanding the means of early diagnostics. The doctoral research project intends to find efficient automated methods of supporting medical doctors in diagnosing renal tumours, which still suffer from insufficient diagnostic methodology. Exploration of tumour representation methods from contrast-enhanced Computed Tomography is the main focus of the research project. The quality of representations is tested in the downstream tasks of abnormality classification into the 8 most prominent subtypes, which are not all visually differentiable. Both deep learning and texture analysis offer comparable efficacy of tumour representation, tested throughout various classification setups. The presented study summarises the project's outcomes, concurrently emphasising more general recommendations for computer vision applications in medical image analysis.