ESTRO 2024 - Abstract Book

S4479

Physics - Machine learning models and clinical applications

ESTRO 2024

Material/Methods:

The proposed pipeline requires 2 image registration networks. The first one warps a patient's MRI to the phase 50 CT (end-exhalation). Subsequently, another network infers the deformation fields which register 4DCT phases to one another. With these networks functioning in conjunction, we warped annotations from a MRI to all instances of the 4DCT (Fig. 1).

Both networks were trained on public datasets. The mono-modal network was validated on public lung CT/CT pairs (209 training, 30 testing) of the NLST dataset [ref.1] from Learn2Reg23.. Meanwhile multi-modal registration was tested on Learn2Reg's public Abdominal MR-CT dataset [ref.2] (8 training + 8 testing MR/CT pairs). Images were resampled to a fixed size of [192,160,192]. Furthermore, to more precisely validate the pipeline, we constructed and annotated an in-house 4DCT/MRI dataset, on which we are currently training : mono-modal registration of different phases from the 4DCT and multi-modal registration from an MRI to a given CT phase. Both networks followed the same LKU-Net architecture [ref. 3]: a UNet with 3 parallel convolution filters of different kernel sizes at each step, and were both trained with the self-supervised approach described by VoxelMorph [ref. 4]. The networks are trained to optimize a loss comprised of (1) a similarity measure between moving/fixed input images (NCC or MI), (2) a regularization penalty and (3) Dice overlap of warped segmentation masks. They output a dense velocity field which is integrated into a smooth displacement field used to align the input images, and thus the anatomical annotations. Finally, we evaluated the approach to this problem by comparing the Dice scores of the annotations before and after registration. The smoothness of the deformation fields (number of foldings + standard deviation of logarithm of the determinant of Jacobian - SDlogJac) and, when available, the Target Registration Error of matched keypoints. The plausibility of the deformations, can also be considered through the scope of the dice metrics which inform us of the alignment of the labeled anatomies.