ESTRO 2024 - Abstract Book

S4480

Physics - Machine learning models and clinical applications

ESTRO 2024

Results:

On the NLST CT/CT validation dataset, the trained network achieves a mean dice score of 0.986 (0.908 before registration), and a post-registration TRE of 1.301 ± 0.121mm (9.703 ± 2.202mm before registration). On the public multi-modal task, the Dice score is improved from 0.509 to 0.859 with very few training samples (8 training samples). Figure 2(a) depicts mean differences before/after registration on several axial slices from a random pair of CTs.

Preliminary results on a small subset of our in-house dataset (10% of samples) indicate a Dice score of 0.951 (0.907 before registration) for registrations between 4DCT phases.

Furthermore these networks output sane deformation fields, with little to no foldings, and smooth deformation transitions: 0.063 ± 0.038 standard deviation of the deformation fields' intra-mask determinant of Jacobian on the NLST data (0.138 ± 0.0317 on the MR/CT dataset), and 0 voxel foldings on the 30 test samples from the NLST data (4.7e-6 on average for the MR/CT data).

Once fully trained, the network took on average 0.093s to register a pair of images on a NVIDIA GeForce RTX 3090.