Image processing and classification of urothelial carcinoma using tissue sample images

Abstract

Bladder cancer is the 6th most common cancer in the world, with 429.000 new incidents reported in 2012, of which urothelial carcinoma is the most common. Prognostic value of the current grading systems are low, with only a significant difference for progression between lowest and highest grade on one system. As a consequence of a recurrence rates of 50-70% and progression to a higher stage in 10-30% of patients, extensive follow-ups are given regularly over several years after first diagnosis. The objective of this thesis is to determine if a local texture analysis can be used as an aid in the prediction of recurrence and progression on patients originally diagnosed with TaT1 urothelial carcinoma. An analysis is done using microscopic tissue samples from 42 patients. Textures are described using local binary pattern and local variance, and features are computed as the chi-squared of the descriptor histograms and predefined models for each prognoses. Local binary pattern achieves approximately 80% correct identification of patients with recurrence, while identification of patients without recurrence are approximately 50%. Suggesting this descriptor can be used to identify patients with recurrence. Prediction using local variance achieve better than random-guessing using a linear normalization of images, but overall results are low. Prognostic value for progression using both descriptors are low, with no clear identification of patients with progression found.