Mittal, ShachiXue, Baide2025-08-012025-08-012025Xue_washington_0250O_28182.pdfhttps://hdl.handle.net/1773/53449Thesis (Master's)--University of Washington, 2025Adaptation and customization of machine learning models in renal pathology remain an obstacle to researchers' rapid development of computational renal pathology and clinician’s implementation of the published machine learning models. The challenge of adapting and customizing renal pathological models is due to a lack of documentation, data disclosure, and open-source code. To bridge the gap between published models and researchers trying to make use of them, we present a general framework for adapting and customizing machine learning models and a working pipeline in the context of renal pathology image analysis. We leveraged the post-transplant renal data from our collaborators and the Omni-Seg model trained with minimum change disease to create the Auto-OSeg pipeline. Auto-OSeg pipeline can automatically predict pathology image data in .ndpi format, retrain models, and produce test accuracies for the retrained models. Auto-OSeg pipeline is repeatable with automation, accessible through detailed documentation, and expandable with functionally separate scripts. We also provided a framework for the general process of adaptation and customization to help organize the process of adapting image segmentation models in general: data relevance assessment, model assessment, customization implementation, and adaptation and retraining assessment. Auto-OSeg pipeline’s retraining and predictions have shown initial signs of accuracy improvements and will provide a strong foundation for future optimizations.application/pdfen-USCC BY-SAAdaptationAutomationMachine learningRenal PathologySegmentationComputer sciencePathologyChemical engineeringThesis