Accelerated annotation of 3D medical images using interactive segmentation

Abstract

This thesis explores the use of interactive segmentation on 3D medical CT scan images. Interactive segmentation allows users to manually guide image division, enhancing the precision of medical imaging analysis. In this study, the efficacy of DeepEdit model, which is an interactive segmentation mode, for spleen segmentation in CT scans across various medical imaging datasets was explored. The research focuses on the impact of number of guidance points on model performance, illustrating a variable influence depending on dataset complexity and model proficiency. Results suggest that for simpler datasets, such as the Medical Segmentation Decathlon data, additional guidance points only marginally improve the mean Dice Score. In contrast, complex datasets like MVision spleen dataset display significant performance improvement with an increased number of guidance points when a pre-trained model is utilized. However, a model trained specifically on MVision data showed minimal Dice Score enhancement with additional guidance points, indicating a high level of efficiency already achieved. These findings provide a robust foundation for further exploration into the optimization of interactive segmentation models in the context of medical imaging.