ENHANCING MODEL PERFORMANCE IN HIP JOINT SEGMENTATION BY LEVERAGING MULTIPLE IMAGE OUTPUTS FROM DIXON MRI

Abstract

INTRODUCTION

Exploring ways to improve methodologies for model training in MRI-based segmentation tasks is important for enhancing segmentation accuracy. Accurate automated segmentations are essential for applications including the extraction of imaging biomarkers. Leveraging advanced deep learning techniques, such as nnU-Net, holds promise for achieving precise segmentations. Additionally, incorporating multiple image outputs from Dixon MRI into these methodologies could improve segmentation performance.

OBJECTIVE

The aim of the study was to investigate whether the accuracy of an automated deep learning segmentation model for the hip joint could be improved by combining information from multiple image outputs from a Dixon sequence.

METHODS

Manual segmentations of 20 participants (40 hips) of the Generation R study, comprising individuals who were 19 years old (11 males, 9 females), were used to train an nnU-Net. The segmented regions included the femoral bone, acetabular bone, femoral cartilage, and acetabular cartilage in both the left and right hip. The 2D and 3D configurations of the nnU-Net were trained using 5-fold cross-validation and an ensemble of the 2D-3D configurations was used in the analyses. The input consisted of 1) water-only images from the Lava Flex sequence, with an in-plane resolution of 0.89 × 0.89 mm² and a slice thickness of 1.2 mm, 2) a combination of in-phase and water-only images, as well as 3) a combination of water-only and fat-only images from the same sequence, allowing for comparison between the methodologies. The combination of water-only and in-phase images was chosen for potentially increasing segmentation accuracy, following visual inspection. Additionally, the water-only and fat-only combination aimed to assess if segmentation results could be enhanced by offering additional tissue contrast. Evaluation was performed on a hold-out test set consisting of 10 manually segmented hips, using the Dice similarity coefficient (DSC) and mean surface distance (MSD).

RESULTS

The mean ±SD DSC for the segmentation of bone and cartilage in the hip using the 1) water-only vs 2) in-phase and water-only combination vs. 3) combination of fat-only and water-only images were as follows: femoral bone 1) 0.961±0.013 vs. 2) 0.967±0.004 vs. 3) 0.950±0.036, acetabular bone 1) 0.886±0.027 vs. 2) 0.893±0.021 vs. 3) 0.896±0.016, femoral cartilage 1) 0.762±0.018 vs. 2) 0.763±0.020 vs. 3) 0.752±0.023, and acetabular cartilage 1) 0.760±0.028 vs. 2) 0.768±0.022 vs. 3) 0.768±0.022. The MSDs were: femoral bone 1) 0.412± 0.105 mm vs. 2) 0.341±0.047 mm vs. 3) 0.297±5.59 mm, acetabular bone 1) 0.602±0.109 mm vs. 2) 0.567±0.097 mm vs. 3) 0.651±0.077 mm, femoral cartilage 1) 0.276±0.033 mm vs. 2) 0.271±0.034 mm vs. 3) 0.295±0.033 mm, and acetabular cartilage 1) 0.313±0.078 mm vs. 2) 0.282±0.049 mm vs. 3) 0.281±0.046 mm.

CONCLUSION

The results indicate that combining the in-phase and the water-only images for model training improves segmentation quality. Specifically, improved overlap and better alignment of the boundaries was observed, as indicated by higher DSCs, and reduced MSD. These results underscore the efficacy of using combined image outputs to enhance the accuracy of hip segmentation in MRI.