Learning Generalizable Features for Tibial Plateau Fracture Segmentation Using Masked Autoencoder and Limited...

Area of Science:

• Medical imaging and artificial intelligence
• Orthopedic surgery and radiology

Background:

• Accurate segmentation of tibial plateau fractures (TPF) in computed tomography (CT) is crucial for diagnosis and treatment planning.
• Deep learning models for TPF segmentation require large annotated datasets, which are costly and time-consuming to acquire due to expert knowledge demands.
• Existing semi-supervised methods struggle with the complexity of fracture patterns and limited generalizability.

Purpose of the Study:

• To develop an effective training strategy for accurate TPF segmentation in CT using Masked Autoencoder (MAE) pretraining.
• To reduce the reliance on extensive manual annotations for deep learning models.
• To enhance the generalizability and transferability of TPF segmentation models.

Main Methods:

• Proposed a training strategy utilizing MAE pretraining on unlabeled CT data to capture skeletal structures and fracture details.
• Fine-tuned the MAE-pretrained model using a small subset of labeled TPF CT scans.
• Evaluated the method on an in-house dataset (180 CT scans) and a public pelvic CT dataset.

Main Results:

• Achieved high segmentation accuracy with an average Dice Similarity Coefficient (DSC) of 95.81%, average Symmetric Surface Distance (ASSD) of 1.91mm, and Hausdorff distance (95HD) of 9.42mm using only 20 annotated cases.
• Outperformed existing semi-supervised methods in TPF segmentation.
• Demonstrated strong transferability to hip fracture segmentation on a public pelvic CT dataset.

Conclusions:

• The proposed MAE-based training strategy effectively reduces annotation requirements for accurate TPF segmentation in CT.
• The method enhances model performance and generalizability, offering a practical solution for clinical applications.
• The approach shows potential for broader applications in medical image segmentation tasks.