Transfer learning-based approach for automated kidney segmentation on multiparametric MRI sequences.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Multiparametric magnetic resonance imaging (mp-MRI) offers superior soft tissue contrast for kidney cancer detection.
• Developing supervised kidney segmentation algorithms is challenging due to the need for manual labels for each mp-MRI protocol.

Purpose of the Study:

• To develop a transfer learning approach for automated kidney segmentation on limited mp-MRI datasets.
• To improve the accuracy and efficiency of kidney segmentation across various mp-MRI sequences.

Main Methods:

• A 2D attention U-Net model was developed for kidney segmentation on T1-weighted nephrographic phase contrast-enhanced (CE)-MRI (T1W-NG).
• Pretrained weights from the T1W-NG model were transferred and fine-tuned to segment kidneys on five other mp-MRI sequences (T2W, T1W-IP, T1W-OP, T1W-PRE, T1W-CM).
• Model performance was evaluated using Dice Similarity Coefficient (DSC), absolute volume difference, Hausdorff distance, and center-of-mass distance via fivefold cross-validation.

Main Results:

• The T1W-NG model achieved a kidney segmentation DSC of 90.5%.
• Transfer learning models showed an average DSC increase of 2.96% across the five additional mp-MRI sequences compared to randomly initialized models.
• Specific DSC improvements included: T2W (87.19% to 89.90%), T1W-IP (83.64% to 85.42%), T1W-OP (79.35% to 83.66%), T1W-PRE (82.05% to 85.94%), and T1W-CM (85.65% to 87.64%).

Conclusions:

• A pretrained model for automated kidney segmentation on one mp-MRI sequence effectively enhances segmentation performance on other sequences.
• Transfer learning reduces the dependency on large, manually labeled datasets for developing robust kidney segmentation algorithms.
• This approach shows significant potential for improving automated kidney segmentation in clinical settings using diverse mp-MRI protocols.