An AI-Based Approach to Inter-Subject Deformable Image Registration
J Jiang*, M Thor, A Rimner, J Deasy, H Veeraraghavan, Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
MO-I430-BReP-F3-1 (Monday, 7/11/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 3
Purpose: A quantitative understanding of radiotherapy toxicity is elusive. Understanding the location of sensitive tissues that drive radiotherapy toxicity often requires accurately registering diverse patients and dose distributions to a single reference anatomy. However, this is difficult with current methods as patients can differ significantly in anatomy including different tumor characteristics. To address this, we developed a novel tumor-aware recurrent deep learning deformable image registration (DIR) method (TRec) for aligning CT scans of patients with lung cancers that is robust to the presence of tumors.
Methods: A recurrent deep learning DIR method with tumor mass deformation constraints was developed on internal (N = 204) and evaluated on two external datasets using 1,136 CT image volume pairs of patients diagnosed with lung cancer. In order to model large deformations, a 3D convolutional long short term memory (CLSTM) network architecture was used that computes a progressive dense deformation flow field that operates by incrementally improving regional alignments. A tumor mass constraint was introduce to penalize deviations of local Jacobian determinants of deformation from unity, resulting in tumor mass preservation. The network was trained in an unsupervised way without requiring ground truth deformation vector fields. The resulting registration accuracy was tested with expert segmented tumors (N = 756 pairs; dataset I) as well as deep learning auto-segmented tumors (N = 380 pairs; dataset II).
Results: TRec produced significantly (p < 0.001) more accurate registrations than standard DIR methods and other deep learning DIR methods. The average Dice similarity coefficient (DSC) quantifying the registration-based tumor segmentation accuracy was 0.81 which was comparable to 5 raters (DSC=0.82). TRec achieved a DSC of 0.81,0.93,0.86,0.70 for tumors, lungs, heart, and spinal cord.
Conclusion: A novel deep learning based method to perform inter-patient CT scan DIRs for images containing tumors was developed and validated on two independent datasets.
Funding Support, Disclosures, and Conflict of Interest: Andreas received grants from Varian Medical Systems, AstraZeneca,Merck, Pfizer, Boehringer Ingelheim.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email