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

TL;DR

The paper tackles automated tibial plateau fracture segmentation from CT where annotated data are scarce due to annotation difficulty. It introduces a pretrain–finetune pipeline using Masked Autoencoder (MAE) pretraining on unlabeled CT scans to learn anatomical priors and fracture cues, followed by fine-tuning with a small labeled set using a UNETR-based decoder. Experiments on an in-house dataset of 180 scans and a public pelvic-fracture dataset demonstrate superior performance over semi-supervised baselines and strong cross-dataset transferability. The approach reduces annotation burden and offers a scalable path for robust fracture segmentation across diverse datasets.

Abstract

Accurate automated segmentation of tibial plateau fractures (TPF) from computed tomography (CT) requires large amounts of annotated data to train deep learning models, but obtaining such annotations presents unique challenges. The process demands expert knowledge to identify diverse fracture patterns, assess severity, and account for individual anatomical variations, making the annotation process highly time-consuming and expensive. Although semi-supervised learning methods can utilize unlabeled data, existing approaches often struggle with the complexity and variability of fracture morphologies, as well as limited generalizability across datasets. To tackle these issues, we propose an effective training strategy based on masked autoencoder (MAE) for the accurate TPF segmentation in CT. Our method leverages MAE pretraining to capture global skeletal structures and fine-grained fracture details from unlabeled data, followed by fine-tuning with a small set of labeled data. This strategy reduces the dependence on extensive annotations while enhancing the model's ability to learn generalizable and transferable features. The proposed method is evaluated on an in-house dataset containing 180 CT scans with TPF. Experimental results demonstrate that our method consistently outperforms semi-supervised methods, achieving an average Dice similarity coefficient (DSC) of 95.81%, average symmetric surface distance (ASSD) of 1.91mm, and Hausdorff distance (95HD) of 9.42mm with only 20 annotated cases. Moreover, our method exhibits strong transferability when applying to another public pelvic CT dataset with hip fractures, highlighting its potential for broader applications in fracture segmentation tasks.

Figures (6)

• Figure 1: Illustration of tibial plateau fractures (TPF) in computed tomography (CT) slices from three different views: (a) axial view, (b) sagittal view, and (c) coronal view. The fractures exhibit considerable variability in location and morphology, with some fragments being particularly difficult to distinguish, presenting huge challenges for the accurate segmentation.
• Figure 2: The proposed pretrain and finetune framework for TPF segmentation in CT. Note that the pretrain stage uses unlabeled data while the finetune stage leverages limited labeled data.
• Figure 3: Architecture of the ViT block used in the MAE.
• Figure 4: Architecture of the UNETR.
• Figure 5: 3D visualization of the segmentation results for tibial plateau fractures.