Deep Learning Auto-Segmentation on Multi-Sequence MRI for MR-Guided Adaptive Radiation Therapy

Abstract

Rapid and accurate generation of synthetic CT (sCT) from daily MRI, required in MR-guided adaptive radiotherapy (MRgART), is challenging in abdomen due to the air volumes that can change quickly and randomly (thus, no paired CT available) and hard to automatically segment on daily MRI. This work aims to develop a novel structure-preservation deep learning method to quickly create sCT from a special MRI sequence. Materials/Methods: The sCT model was based on the generative adversarial networks (GANs) technology with innovations including extra deformable layers in sub-networks and mutual information loss terms, which were added to effectively guide the network to preserve true structures from MRI for those highly deformed organs and air pockets in abdomen. The model was developed to use air scan, a specially designed MRI sequence to image air in abdomen, which a 3D FLASH Cartesian sequence that had optimized RF pulses to achieve a minimum echo time of 1 msec with heavy acceleration to keep scan times under 9 seconds to avoid motion artifacts. Daily air scans acquired along with daily plan MRI on a 1.5T MR-Linac during MRgART for 21 patients with abdominal tumors were used to create the sCT model (sCT-DL). The sCT-DL was compared with the results from: (1) sCT-DIR, generated via deformable image registration (DIR) of reference CT and daily plan MRI, and (i) sCT-Gdiff, a previously reported method to automatically segment air volumes on daily MR by a threshold in a union of (i) deformed air-containing organs (e.g., bowels) and (ii) a expansion to account for DIR inaccuracy (calculated by taking the root mean square error between primary and deformed secondary images, divided by the gradient of primary). In addition, the air volumes on sCT-DL obtained with a threshold of HU < -300 and on sCT-Gdiff were compared to those manually delineated on the air scans (Airmanual) based on Dice similarity coefficient (DSC). Dose calculated for a MR-Linac plan on sCT-DL was compared to those calculated on sCT-DIR and sCT-Gdiff. Dosimetric accuracy was measured using the fractional volume with dose disagreement < 3% (FV3). The bone volumes from sCT-DL were compared to those on sCT-DIR (ground truth) based on DSC and FV3. Results: The sCT-DL creation was very fast (< 1.0 sec with a hardware of 28 CPUs and P2000 GPU). The air volume DSC was 0.49 § 0.1 for sCT-DL and 0.89 § 0.06 for sCT-Gdiff, as compared to the Air-manual volumes. For dose accuracy, the FV3 was 0.86 § 0.03 for sCT-DL versus 0.90 § 0.01 for sCT-Gdiff. For the bone volumes on sCT-DL, DSC was 0.54 § 0.04, and FV3 was 0.87 § 0.01 as compared to the ground truth of sCT-DIR. Conclusion: It is promising to use the proposed novel structure-preservation deep learning method to automatically generated sCT in abdomen based on this proof-of-principle study. The sCT can be generated within 10 sec including the special MRI acquisition. With further development using large datasets, the novel sCT method may be implemented for MRgART of abdominal tumors.

Automatic Contour Refinement for Deep Learning Auto-segmentation of Complex Organs in MRI-guided Adaptive Radiation Therapy

“…This may explain why a better ACR performance was achieved for the combined bowels compared with the small bowel and large bowel separately, because the adopted CNN models had some difficulty in differentiating the small and large bowels. 30 , 31 …”

Editorial for “A Fully‐Automatic Method to Segment Choroid Plexuses in Multiple Sclerosis Using Conventional MRI Sequences”

“…1,2 These inflammatory changes in MS patients are associated with CP enlargement, which is linked to a poorer pathophysiological progression, higher Expanded Disability Status Scale score, increased relapse rate, and worsening of clinical disability. 3,4 The research community has invested significant efforts in developing and validating imaging segmentation techniques in MS. 5,6 Despite manual segmentation being considered the gold standard for CP segmentation, it poses several challenges, including time-consuming procedures and intra-and inter-rater variability. 7 Notably, developing automatic methods for CP segmentation is crucial to study these structures on large cohorts of subjects.…”

“…The research community has invested significant efforts in developing and validating imaging segmentation techniques in MS 5,6 . Despite manual segmentation being considered the gold standard for CP segmentation, it poses several challenges, including time‐consuming procedures and intra‐ and inter‐rater variability 7 …”