Introduction: Medical image segmentation is used daily in clinical practice. Although manual segmentation is still considered the gold standard, it is very time consuming. However, with the development of deep learning, semi-automated and automated segmentation models are gaining popularity as an efficient alternative to manual segmentation. One of the medical image segmentation software is 3D Slicer, which is a free open-source software application. It also offers various extensions such as Monai Label, which is an intelligent tool for automatic medical image segmentation based on deep learning methods. Purpose: The aim of the master thesis is to establish Monai Label web application, integrate it with 3D Slicer and test its performance on downloaded CT scans. We compared the semi-automatic segmentation procedures of 3D Slicer with the automatic segmentation procedures of the Monai Label extension on CT studies acquired from the TCIA online database, where we segmented the liver, kidneys and bladder. Methods: In this study, we demonstrated the establishment of the Monai Label web application and the establishment of communication with the 3D Slicer program. Furthermore, we performed segmentations of the liver, kidneys, and bladder on the downloaded CT studies using a semi-automatic region-growing segmentation procedure and an automatic segmentation procedure in the Monai Label extension, using the whole_body_ct_segmentation model. To evaluate the effectiveness of the segmentation procedures, we performed a Dice calculation using segmentations from the TCIA online database, more precisely the CT-ORG image database, as reference segmentations. Results: Analysis of the results in our research confirms the high efficiency of both semi-automatic and automatic segmentation methods for liver, kidneys and bladder. The semi-automatic segmentation of liver and kidneys achieved average Dice coefficients of 0,95 and 0,91, which is comparable to automatic segmentations, where we achieved Dice coefficients of 0,96 for liver and 0,94 for kidneys with the whole_body_ct_segmentation model. Bladder segmentation showed similarly high performance, with average Dice coefficients of 0,92 for both semi-automatic and automatic methods. Automatic segmentations were performed significantly faster, in an average of 1,64 minutes for 104 organs, which represents 1116 times faster performance compared to semi-automatic methods. Discussion and conclusion: 3D Slicer proved to be easy to use from inexperienced user standpoint, and integration with the Monai Label extension was also effortless. Evaluation of the methods showed that automatic models such as whole_body_ct_segmentation achieve high Dice efficiencies and incomparable time savings compared to semi-automatic segmentations. The applications of segmentation methods are wide ranging, from treatment planning to diagnostics and education to research, where automated segmentations allow for more accurate and faster procedures. Despite the advantages of automatic methods, the quality of the models and the appropriate settings need to be considered for optimal results, especially when used in clinical settings.
|
