2207 - Re-Evaluating the Necessity of Pre-Processing Magnetic Resonance Images for Deep Learning-Based Organ Segmentation: A Study on Segmentation Challenges Using MR-Linac Images

A. Amjad, C. Sarosiek, Y. Zhang, R. Conlin, M. Zarenia, B. A. Erickson, W. A. Hall, and E. S. Paulson; Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI

Purpose/Objective(s): Magnetic resonance images (MRIs) often undergo pre-processing, including bias field removal, filtering, and histogram matching or intensity normalization prior to use in radiation therapy. While pre-processing generally improves the consistency of MR images across field strengths and vendors, with sufficient training data and appropriate algorithm deep learning auto-segmentation (DLAS) may not require pre-processing to produce accurate segmentation. We investigate here the impact of pre-processing on auto-segmentation accuracy using four DLAS models for abdominal MR images obtained during MR-guided online adaptive radiotherapy (MRgOART) of abdominal tumors. Materials/

Methods: In this study a series of DLAS models were developed to contour abdominal organs on non-fat suppressed images acquired on the 1.5T MR-Linac. DLAS model-A was developed using MRIs that were pre-processed with bias field correction (N4 algorithm), noise filtering using anisotropic diffusion, and intensity normalization using volumetric median. DLAS model-B was developed with pre-processed MRIs, and additionally the slope, intercept and pixel representation were fixed at 1, 0, and 1, to remove any negative intensities, and standardize intensities across the multiple sequences. DLAS models (C1, C2) were developed without pre-processed MRIs. DLAS models-A, B, and C1 were developed by one vendor, while DLAS model-C2 was developed by a second vendor. All models were trained using 107 datasets obtained from a 1.5T MR-Linac and tested on the same 21 datasets. Qualitative checks and five metrics were used to evaluate the accuracy of the segmentation of the complex abdominal organs.
Results: Negative intensity values were observed when pre-processed images were saved to disk prior to use in DLAS, model-A. Average values of the 3D metrics across the 5 organs are shown in the table below. Model-C1 demonstrated higher accuracy, with mean DSC of 0.7, 0.93, 0.94, 0.7, 0.9, for duodenum, left kidney, liver, pancreas, and stomach. The best segmented large bowel was seen with model-B, with 0.8 DSC, whereas model-A had the lowest accuracy, DSC of 0.6, 0.5, 0.8, for duodenum, pancreas, and stomach. Model-C2 performed comparable to model-C1, and better for large bowel, duodenum, and pancreas.
Conclusion: Pre-processing can improve the consistency of MR images. While negative intensities can result when pre-processed images are written to disk prior to use in DLAS algorithms, they have little effect on the accuracy of DLAS using the models trained. However, our study shows that pre-processing may not be required to produce accurate abdominal OAR segmentation on MR-Linac images. Abstract 2207 – Table 1

	Dice Similarity Coefficient DSC	Surface DSC	Relative Added Path Length rAPL	Mean Distance to Agreement MDA	95 percentile Hausdorff Distance 95HD
Model-A	0.74	0.53	30.12	6.17	22.35
Model-B	0.79	0.63	34.33	4.52	17.87
Model-C1	0.76	0.59	14.23	10.02	47.16
Model-C2	0.80	0.65	9.98	6.56	28.57