Cross-modality deep learning: Contouring of MRI data from annotated CT data only.

Area of Science:

• Medical physics
• Radiotherapy
• Artificial intelligence in medicine

Background:

• Manual segmentation of organs-at-risk and targets in radiotherapy is time-consuming and subjective.
• Deep learning for auto-segmentation requires large annotated datasets, which are scarce for MRI in radiotherapy.
• Developing reliable auto-segmentation algorithms is crucial for online adaptive radiotherapy.

Purpose of the Study:

• To develop a method for generating synthetic MR images from annotated CT images.
• To train a convolutional neural network (CNN) for parotid gland segmentation using these synthetic MR images.
• To address the challenge of limited annotated MRI data for radiotherapy.

Main Methods:

• Utilized a 2D CycleGAN network to perform cross-modality image translation from CT to MRI.
• Propagated contours from CT images to generate annotations for synthetic MR images.
• Trained a 2D CNN on synthetic MR images and evaluated segmentation accuracy on real MR images using DSC, HD, and MSD metrics.

Main Results:

• Achieved segmentation accuracy (DSC: 0.77±0.07) close to interobserver variation (DSC: 0.84±0.06).
• Segmentation performance was comparable to training a CNN directly on CT images (DSC: 0.81±0.07).
• Quantified accuracy using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and Mean Surface Distance (MSD).

Conclusions:

• Cross-modality learning effectively addresses sparse training data challenges in medical image segmentation.
• The method enables the generation of annotated synthetic MR images from readily available annotated CT datasets.
• This technique facilitates adaptation of segmentation models across different imaging modalities and MRI contrasts.