An MRI-Based Global Deep Learning Auto-Segmentation Model for Abdominal Organs

Abstract

<h3>Purpose/Objective(s)</h3> Multi-sequence MRIs are often used in MRI-based radiation treatment (RT) planning and/or MR-guided adaptive RT (MRgART). As MRI sequences are constantly evolving to reflect changing technology and/or tumor specific RT imaging, it is desirable to develop sequence independent deep learning auto-segmentation (DLAS). In this study, we present a global DLAS model that can be used for segmenting abdominal organs on multi-contrast images from multiple machines including those different from training. <h3>Materials/Methods</h3> The global DLAS model was built on a previously investigated 3D neural network for DLAS with multi-site CTs and multi-sequence MRIs. 167 abdominal MRIs acquired with 5 unique sequences with a wide range of acquisition parameters: T2-HASTE and T1-DIXON, fast field echo (FFE), balanced turbo (BTFE) and TFE daily motion-averaged MRI, acquired from a 3T MR simulator and a 1.5T MR-Linac, were used to train the global DLAS model. An additional 81 datasets, including 49 sets of varying image quality and contrast, different to those in the training datasets, were used for testing. The model performance was evaluated by comparing DLAS with manually drawn ground truths, for 10 abdominal organs using 4 metrics: Dice similarity coefficient (DSC), surface DSC (sDSC), mean distance to agreement (MDA), and relative added path length (APL_rel), defined as the APL divided by the organ volume. <h3>Results</h3> The global DLAS model performed well for organs with stable shapes and minimum volume variations (e.g., aorta, liver, kidneys, and spleen) with average DSC = 0.82-0.98, sDSC = 0.67-1, MDA = 0.5-2.2 mm, and APL_rel = 0.1-12 mm/cc, showing considerable invariance to acquisition parameters. High accuracy was also observed for organs with consistent anatomical changes (stomach and large bowel), with average DSC = 0.77-0.94, sDSC = 0.51-0.82, MDA = 0.43-6.3 mm, and APL_rel = 5-18.3 mm/cc. For small bowel, duodenum, and pancreas, reasonable accuracy on the 3T MR simulator images with average DSC = 0.74-0.85, sDSC = 0.65-0.74, MDA = 1.85-3 mm, and APL_rel = 11-21 mm/cc, indicated improved performance for images with higher soft tissue contrast and less motion artifacts. Interestingly accurate segmentation was observed in 23 out of 49 sequences different from training datasets. For the images with relatively poorer quality (e.g., motion artifacts and low signal density in MR-Linac images) and/or distinct contrast (e.g., fat suppressed), the DLAS performance was relatively poorer, indicating inclusion of larger datasets of these images to the model training. <h3>Conclusion</h3> It is feasible to develop an MRI-based, sequence independent, global DLAS model using MRI datasets of limited sequences and assess how well the model output extends to sequences different from training, with variable contrast and image quality. With further improvement using larger training datasets, the global model could be useful, particularly for MRgART where MRI sequences may be updated or adjusted routinely or frequently.