Deep learning application for abdominal organs segmentation on 0.35 T MR-Linac images

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Radiation oncology

Background:

• Magnetic resonance (MR) imaging-guided linear accelerators (linacs) offer enhanced soft tissue contrast for abdominal radiation therapy.
• Accurate segmentation of abdominal organs at risk (OARs) is crucial for treatment planning but is currently a manual, time-consuming process with inter-observer variability.

Purpose of the Study:

• To investigate deep learning-based automatic segmentation solutions for abdominal OARs using 0.35 T MR images.
• To evaluate the performance of various UNet-based models and identify the optimal approach for clinical application.

Main Methods:

• One hundred and twenty-one sets of abdominal MR images and ground truth segmentations were collected.
• Several 2D UNet-based models (Classical UNet, ResAttention UNet, EfficientNet UNet, nnUNet) were trained, with the best performing model subsequently trained using a 3D strategy.
• Performance was evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD), and volume analysis.

Main Results:

• The 3D nnUNet model achieved the highest performance, with DSC scores of 0.96 ± 0.01 for the liver, 0.91 ± 0.02 for the kidneys, and 0.91 ± 0.01 for the spinal cord.
• Good IoU scores were observed for these organs, while segmentation for the stomach and duodenum showed lower but still potentially applicable results.
• Hausdorff Distance and volume analysis corroborated the segmentation accuracy, with some variability for the duodenum.

Conclusions:

• The 3D nnUNet model demonstrates significant potential for accurate and efficient automatic segmentation of abdominal OARs in 0.35 T MR-Linac images.
• While duodenum segmentation requires further optimization, the overall results suggest a viable clinical application for improving radiation therapy planning and reducing manual segmentation burdens.