1. What is the aim of developing a generic multi-sequence DLAS model for abdominal organs based on MRI sequences?
The aim of developing a generic multi-sequence DLAS model for abdominal organs based on MRI sequences is to improve efficiency in MRI-based radiation therapy (RT) planning and delivery. The model aims to take full advantage of information from multiple MRI sequences and improve organs at risk (OARs) and tumor segmentation. By developing a single DLAS model for multi-sequence MRIs, the process of manual segmentation of acquired images can be avoided, reducing time and effort. The model also aims to address the challenges associated with the complexity of the abdomen, such as the variability of shape and volume of digestive organs and the occurrence of motion and intensity artifacts in MRI. The generic multi-sequence DLAS model will be evaluated based on its clinical applicability, including the accuracy of the auto-segmented contours and the time-saving benefits compared to manual contouring.
read more
2. What image acquisition parameters are used for MRI datasets?
The image acquisition parameters for MRI datasets include bias field correction using an N4 algorithm, noise filtering using anisotropic diffusion, and intensity normalization by thresholding to volumetric median. These parameters are essential for pre-processing the images to ensure consistency and accuracy in the dataset. The parameters are summarized in Table 1, and representative slices of the 4 image sequences before and after standardization are shown in Figure 1. The Z-score normalization method is used on both T1 and T2 weighted images for each patient base to accommodate variations in the multi-sequence images. Contours of the 12 organs of interest are created or reviewed slice-by-slice by a trained researcher, and manual contouring is done using a commercial clinical contouring tool. These contours are considered as manual reference contours (MRC) in the DLAS training and testing.
read more
3. What data augmentation techniques were used in DLAS model training?
Two data augmentation techniques were adopted in DLAS model training. The first technique involved an in-house developed 3D elastic transformation with minor random deformation applied on images and labels for each case. This method aimed to accelerate DLAS model training by learning embedding features instead of memorizing pixel locations of organs. The second technique employed a gamma intensity transformation with a probability of 0.3, using a gamma uniform distribution ranging from 0.7 to 1.3. This approach mimicked intensity variations across different MRIs, enhancing model robustness and preventing overfitting. These techniques were implemented to create additional training data with larger variations, improving the model's robustness and generalization capabilities.
read more
4. What metrics were used to evaluate auto-segmentation models?
The following quantitative metrics were calculated to evaluate auto-segmentation models: 1. DSC for volumetric overlap, 2. MDA for mean distance between contour sets, 3. HD95% for maximum distance between contour sets, 4. PVD for volume difference, 5. sDSC for surface overlap, and 6. APL/mm for boundary correction. Additionally, rAPL was introduced to calculate the editing distance per organ independent of the organ volume. Tolerance of 2 mm was used for sDSC and APL calculations based on clinical expectations.
read more