United multi-task learning for abdominal contrast-enhanced CT synthesis through joint deformable registration

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Artificial Intelligence in Medicine

Background:

• Synthesizing contrast-enhanced computed tomography (CECT) from non-enhanced CT (NECT) is crucial for radiotherapy target volume delineation, aiming to minimize contrast agent risks and registration errors.
• Current methods often treat image synthesis and registration as separate tasks, overlooking their interdependence and failing to address inter-image misalignment.
• NECT images possess structural information vital for differentiating lesions from surrounding tissues, yet this is not fully leveraged in existing separate-task approaches.

Purpose of the Study:

• To develop a united multi-task learning (UMTL) framework for the joint synthesis and deformable registration of abdominal CECT images.
• To improve the accuracy and reliability of synthetic CECT images for radiotherapy applications.
• To address the limitations of separate synthesis and registration methods by integrating them into a collaborative framework.

Main Methods:

• An end-to-end UMTL framework was designed, integrating a deformation field learning network for misalignment correction and a 3D generator for CECT synthesis.
• Enhanced component images and a multi-loss function were employed to boost the performance of the synthetic CECT images.
• The framework was evaluated on datasets with varying resolutions and a separate external test dataset.

Main Results:

• The proposed method achieved a mean absolute error (MAE) of 32.78±7.27 HU for synthetic venous phase CECT images on the external test dataset.
• Specific to the liver region, the mean MAE was 24.15±5.12 HU, with a peak signal-to-noise ratio (PSNR) of 27.59±2.45 dB and a structural similarity index (SSIM) of 0.96±0.01.
• High Dice similarity coefficients (0.96±0.05 for high-resolution and 0.95±0.07 for low-resolution) were obtained for the liver region between true and synthetic CECT images.

Conclusions:

• The UMTL framework effectively integrates CECT image synthesis and deformable registration, outperforming separate approaches.
• The method demonstrates high accuracy and similarity in synthetic CECT images, particularly in the liver region.
• This approach holds significant potential for enhancing radiotherapy target volume delineation by providing reliable synthetic CECT images.