Deep learning auto-segmentation on multi-sequence magnetic resonance images for upper abdominal organs.

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Multi-sequence MRIs are crucial for defining targets and organs at risk (OAR) in radiation therapy (RT).
• Current deep learning auto-segmentation models primarily use single MRI sequences.
• There is a need for advanced auto-segmentation methods utilizing multi-sequence MRI data.

Purpose of the Study:

• To develop and evaluate a multi-sequence deep learning-based auto-segmentation (mS-DLAS) model.
• To leverage multi-sequence abdominal MRIs for improved auto-segmentation accuracy.
• To enhance organ and target delineation in radiation therapy planning.

Main Methods:

• A 3DResUnet network was trained using 4 T1 and T2 weighted MRI sequences from 71 abdominal tumor cases.
• Data pre-processing, Z-normalization, and data augmentation strategies were implemented.
• Performance was evaluated using metrics like Dice Similarity Coefficient (DSC) and Mean Distance to Agreement (MDA), comparing mS-DLAS with sequence-specific models.

Main Results:

• The mS-DLAS model achieved an average DSC of 0.87 and MDA of 1.79 mm over 12 upper abdominal organs.
• Segmentation of 12 upper abdominal organs was completed within 21 seconds per case.
• The model demonstrated robustness by successfully segmenting MRI sequences not included in its training set.

Conclusions:

• A novel MRI-based mS-DLAS model for auto-segmenting upper abdominal organs has been developed.
• Multi-sequence segmentation is valuable for accurate delineation in clinical RT, especially for abdominal tumors.
• This work advances fast and accurate segmentation on multi-contrast MRI, paving the way for MR-only radiation therapy.