Renal Convolutional Models: Feasibility, Data Representation, and CT Imaging in Renal Cancer Diagnosis

This thesis explores the utilization of convolutional neural networks (CNN) with CT imaging data for a non-invasive diagnosis of renal cancer tumors. The feasibility of this approach is explored through the development and maturation of a CNN model, progressing from proof-of-concept, to alignment with the given deployment environment. Various strategies for enhancing the performance, explainability, and contextual alignment of the CNN are examined. These include experimentation with data representation, encompassing file formats, color representations and data augmentation techniques specific to image data with a focus on renal CT imaging data.

The expected usage of image features, color representation in this context, is also explored revealing a need for analysis of usage context and user expectation. Additionally, exploratory data analysis within a clinical context provides valuable insights into the usage, processing, and representation of renal CT imaging data, further aligning the model with the usage context. Furthermore, the concept of supporting models is introduced to automate interpretation tasks in a clinical deployment setting. This concept emphasizes the importance of understanding and mapping the data pipeline across research and deployment environments, with a focus on transitioning from manual intervention to automation of complex tasks.

In conclusion, this research underscores the feasibility of CNN’s for renal cancer tumor classification. It emphasizes the significance of data exploration, understanding, and processing, highlighting the need for an approach where model development is guided by data representation and alignment to the characteristics of the context. Future research is presented with an emphasis on further exploring the deployment aspect of the model in terms of MLOPS and data processing automation, but also enhancing data representation further through elimination of extraneous feature data.