Impact of structurally constraining loss functions in contour delineation for adaptive radiotherapy

Abstract

Image processing techniques are critically important to adaptive radiotherapy (ART) as they can improve the treatment and help minimise radiation toxicity. Contour delineation is one of the essential processes in ART based cancer treatment as it can tailor the treatment to the anatomy on any particular day, thus providing a better treatment. The aim of this research is to study how a structurally constrained loss function influences contour delineation when it estimates a new treatment plan by establishing a geometrical relationship between the planning scan acquired prior to treatment and a time-of-treatment scan.

A block matching based registration method with two-pass regularisation was developed to automate contour delineation. Using the planning scan and on-the-day-of-treatment scan, this registration effectively restrains unnatural and unrealistic tissue deformations. These scans were axial CT scans acquired from abdominal region for prostate cancer treatment. A twopass distance and neighbourhood-orientation based regularisation applied during the motion vector estimation in the block matching based non-rigid registration was able to perform contour delineation better than parametric (demons), non-parametric (b-spline) and state-ofthe- art (pyramidal block matching) methods. The Dice scores of the b-spline’s, demons’ and pyramidal block matching’s average Dice scores are respectively 7.9%, 10.66% and 3.49% lower than the Dice score of the proposed method. The averaged computational time of disparity-regularised block matching is respectively 2.7, 1.15 and 1.8 times less than the averaged computational time of the b-spline, demons, and pyramidal block matching. The averaged normalised mean square error of the disparity-regularised block matching is respectively 0.69, 1.06 and 3.75 times less than the averaged normalised mean square error of the b-spline, demons, and pyramidal block matching.

The disparity-regularised block matching was modified to the Lung Computed Tomography (CT) image analysis. The Lung CT usually has large coarser regions and disparity-regularised block matching uses mean absolute error as the image similarity metric which does not consider image texture. In order to include the texture information in the block matching criteria, texture feature maps were used as a part of the block matching process along with the voxel intensities and the displacement vector field estimation was evaluated using the image quality metrics namely, structural similarity index (SSIM) and normalised mean squared error. The averaged structural similarity index and the averaged normalised mean squared error between the scans were 0.9960 and 0.4 respectively, showing that the estimated motion vectors and displacement vector fields were closer to the ground-truth geometrical differences between the Lung CT scans.

To facilitate the contour delineation process, deep learning based generative adversarial networks were developed to generate synthetic CT and cone beam CT (CBCT) images. The generative adversarial networks were trained as a part of the CycleGAN with a structural constraint loss to preserve the anatomical structures in the synthetic images. The variants of the structural similarity index metric (SSIM) were included as a training loss in the 2D CycleGAN and the performance of the image synthesis was evaluated using the image quality metrics such as structural similarity index, mean square error, and peak signal-to-noise ratio.

The other variants of the SSIM are multi-scale SSIM, 4-components weighted, 4-components gradient weighted, 4-components weighted multi-scale and 4-components gradient weighted multi-scale SSIM. From the metrics, it was observed that the 2D CycleGAN with 4-gradient SSIM has generated synthetic images with two times better SSIM metrics than the other 2D CycleGAN variants. On the other hand, the 2D CycleGAN with 4-components gradient weighted SSIM generated synthetic image whose mean squared error was 14% lower than the mean squared error of other 2D CycleGAN variants. Also, both these variants generated synthetic images with better peak signal-to-noise ratio than the rest. Similar to the 2D CycleGAN variants, the performance metrics of the 3D variants showed that the CycleGAN with 4-components gradient weighted SSIM generated better synthetic CT and CBCT images.

Finally, to understand the influence of the structural constraint loss in the contour propagation, the 2D and 3D generative adversarial networks were merged with the two-pass regularised block matching rigid registration. The contour propagation by this framework was then assessed using clinical validation metrics and image quality metrics such as Dice score, mean squared error and Hausdorff distance. Finally, in the block matching based non-rigid registration, two-pass distance and orientation regularisation is a type of structural constraint applied in the block matching process. In Image-to-Image synthesis by generative adversarial networks, the inclusion of structural similarity index metrics in the training loss functions is also a type of structural constraint applied in the pseudo image generation process. Overall, this research contributes in understanding the positive impact of structural constraints in traditional and state-of-the-art medical image processing techniques for Adaptive Radiotherapy.