Introduction: Segmentation is a key part of medical image analysis, as it allows images to be divided into multiple segments or meaningful structures, facilitating further processing and interpretation. 3D Slicer is free, open-source software designed for visualising, processing, and segmenting medical images. Its extensions include MONAI Label, MONAI Auto3DSeg, and TotalSegmentator, which are designed for automatic segmentation using artificial intelligence. Purpose: The purpose of this master's thesis is to emphasise the importance of segmentation in radiology and to compare semi-automatic and automatic segmentation methods for pulmonary vessels and the trachea, as offered by the 3D Slicer software. Methods: Descriptive and experimental research methods were used. The first part of the thesis is based on a review of the existing literature, while the second part is based on the practical implementation of segmentation. To carry out this work, we installed the freely available 3D Slicer software on a workstation. We added the MONAI Label, MONAI Auto3DSeg, and TotalSegmentator extensions. Reference segmentations were performed using the semi-automatic Grow from Seeds method in the 3D Slicer program, while other segmentations were performed using the aforementioned models. We evaluated the effectiveness of the obtained segmentations by calculating the Dice coefficient. Results: Automatic models in the 3D Slicer software enable significantly faster segmentation compared to the semi-automatic method of enlarging areas. The median time for semi-automatic segmentation of the pulmonary vasculature was 51.0 minutes [42.8; 55.5], while the median time for segmentation of the trachea was 31.0 minutes [26.3; 35.0]. The models performed automatic segmentation in a few minutes, reducing the segmentation time by a factor of 2 to 16 times, depending on the model. The Dice coefficient values for pulmonary vein segmentation varied considerably. The most successful models were TotalSegmentator (median 0.80 [0.77; 0.81]) and MONAI Label – wholeBody (median 0.83 [0.72; 0.86]). All models achieved good Dice coefficients in trachea segmentation. The best results were achieved by the Auto3DSeg Lungs model (median 0.84 [0.82; 0.86]) and the Auto3DSeg Mediastinal Anatomy TS2 model (median 0.83 [0.81; 0.85]). Discussion and conclusion: 3D Slicer is a reliable, fast, and accessible tool for supporting the processing and labelling of respiratory anatomical structures. Trachea segmentation was accurate and relatively simple, while pulmonary vasculature segmentation still requires further model optimisation. The use of automatic segmentation models can significantly save radiologists time; however, the optimal approach in clinical practice remains a combination of automatic segmentation and expert radiologist validation.
|
