Cross-modality (CT-MRI) prior augmented deep learning for robust lung tumor segmentation from small MR datasets

Area of Science:

• Radiotherapy
• Medical Imaging
• Deep Learning

Background:

• Accurate tumor segmentation in magnetic resonance (MR) imaging is crucial for radiotherapy planning.
• Training deep learning models for MR segmentation is hindered by the scarcity of large, expert-annotated MR datasets.

Purpose of the Study:

• To develop a cross-modality (MR-CT) deep learning segmentation approach that augments training data.
• To generate pseudo MR images from expert-segmented CT images for enhanced model training.

Main Methods:

• A generative adversarial deep learning network was trained to translate CT images to pseudo T2-weighted (T2w) MR images.
• 377 expert-segmented non-small cell lung cancer CT scans were translated into pseudo MR images to augment the training set.
• The method was benchmarked against shallow learning and state-of-the-art adversarial learning-based cross-modality augmentation techniques.

Main Results:

• The proposed approach demonstrated the lowest statistical variability between pseudo and T2w MR images (Kullback-Leibler divergence of 0.069).
• It achieved the highest segmentation accuracy (Dice similarity coefficient of 0.75 ± 0.12) and lowest Hausdorff distance (9.36 mm ± 6.00 mm) on the test dataset.
• Tumor growth estimations using this method were highly similar to expert assessments (P = 0.37).

Conclusions:

• A novel deep learning approach for MR segmentation was developed, overcoming limitations of small datasets by leveraging cross-modality information.
• The method effectively augments segmentation training data by incorporating tumor knowledge into modality translation.
• The results confirm the feasibility and superiority of this approach over existing state-of-the-art methods.