The goal of the thesis is to reimplement a machine-learning digital pathology pipeline, which will compose together several specific and fragmented pipelines developed within RationAI research group (led by assoc. profs. Petr Holub and Tomáš Brázdil). These specific pipelines include, inter alia, training of deep convolutional neural networks for detection of prostate cancer, pipeline for registration of re-stained biological material and training of detection of epithelial tissues, and pipelines focused on explaining behavior of trained deep learning models.

• an analysis of these specific pipelines
• describe their features
• describe their functional inter-dependencies
• provide a design of modular pipeline which can integrate all the requested features

The pipeline has to be flexible to integrate other image processing components, and to handle varying structure of the data. It must also be capable of producing detailed logs regarding the model configuration and performance during training, validation and evaluation. The logs must contain enough information to ensure a reproducibility of the results. Practical implementation has to support at least TensorFlow.