The development of an interactive AI algorithm for the image segmentation of clinical CT datasets

This thesis presents the development and evaluation of an organ-specific interactive AI algorithm designed for the segmentation of clinical CT datasets, specifically focusing on lung CT scans for demonstration purposes. The work addresses the challenge of automating medical image segmentation, a process traditionally performed manually by experts, which is both time-consuming and prone to variability. To overcome these limitations, the thesis proposes a semi-automatic approach using a 2D U-Net model that incorporates user interactions though the second channel of the CT image input. The model allows users to refine automatic segmentations by providing feedback through simple mouse clicks, iteratively improving the segmentation results. Extensive testing on a lung CT dataset demonstrated the model’s ability to achieve high accuracy segmentations in both described attempts, with performance metrics such as the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) showing strong results across both training and test sets. One of the two reported training schedules showed remarkable interpretative capability in correctly incorporating users annotations. Despite some challenges in handling complex intensity value cases, the interactive model enhances user control while maintaining efficiency, suggesting its potential as a practical tool in clinical workflows. Further research is recommended to improve the model’s robustness and interactivity by the optimization of both the training strategy and the training schedule, including refining the custom loss function weights, the optimizers and the learning rates. Exploring more aggressive augmentation strategies, particularly through the use of larger morphological kernels, could also improve the model’s ability to interactively refine segmentation boundaries.