Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Performance Analysis of Deep Learning Models for Femur Segmentation in MRI Scan

Mengyuan Liu, Yixiao Chen, Anning Tian, Xinmeng Wu, Mozhi Shen, Tianchou Gong, Jeongkyu Lee

TL;DR

The paper evaluates four segmentation architectures (U-Net, Attention U-Net, U-KAN, and SAM 2) for femur segmentation in MRI, using 11,164 annotated scans. Under a unified training regime and an ensemble approach, Attention U-Net achieves the top overall Dice score, while U-KAN excels in small ROI regions such as the femoral shaft; SAM 2 performs competitively but does not surpass CNN-based models in most regions. The findings highlight the continuing strength of CNN-based methods for precise bone segmentation in MRI and underscore the value of attention and KAN enhancements, with dataset size and region size significantly influencing model performance. The results have practical implications for automated femur modeling in orthopedics and rehabilitation, guiding model selection and prompting further exploration with larger datasets.

Abstract

Convolutional neural networks like U-Net excel in medical image segmentation, while attention mechanisms and KAN enhance feature extraction. Meta's SAM 2 uses Vision Transformers for prompt-based segmentation without fine-tuning. However, biases in these models impact generalization with limited data. In this study, we systematically evaluate and compare the performance of three CNN-based models, i.e., U-Net, Attention U-Net, and U-KAN, and one transformer-based model, i.e., SAM 2 for segmenting femur bone structures in MRI scan. The dataset comprises 11,164 MRI scans with detailed annotations of femoral regions. Performance is assessed using the Dice Similarity Coefficient, which ranges from 0.932 to 0.954. Attention U-Net achieves the highest overall scores, while U-KAN demonstrated superior performance in anatomical regions with a smaller region of interest, leveraging its enhanced learning capacity to improve segmentation accuracy.

Performance Analysis of Deep Learning Models for Femur Segmentation in MRI Scan

TL;DR

The paper evaluates four segmentation architectures (U-Net, Attention U-Net, U-KAN, and SAM 2) for femur segmentation in MRI, using 11,164 annotated scans. Under a unified training regime and an ensemble approach, Attention U-Net achieves the top overall Dice score, while U-KAN excels in small ROI regions such as the femoral shaft; SAM 2 performs competitively but does not surpass CNN-based models in most regions. The findings highlight the continuing strength of CNN-based methods for precise bone segmentation in MRI and underscore the value of attention and KAN enhancements, with dataset size and region size significantly influencing model performance. The results have practical implications for automated femur modeling in orthopedics and rehabilitation, guiding model selection and prompting further exploration with larger datasets.

Abstract

Convolutional neural networks like U-Net excel in medical image segmentation, while attention mechanisms and KAN enhance feature extraction. Meta's SAM 2 uses Vision Transformers for prompt-based segmentation without fine-tuning. However, biases in these models impact generalization with limited data. In this study, we systematically evaluate and compare the performance of three CNN-based models, i.e., U-Net, Attention U-Net, and U-KAN, and one transformer-based model, i.e., SAM 2 for segmenting femur bone structures in MRI scan. The dataset comprises 11,164 MRI scans with detailed annotations of femoral regions. Performance is assessed using the Dice Similarity Coefficient, which ranges from 0.932 to 0.954. Attention U-Net achieves the highest overall scores, while U-KAN demonstrated superior performance in anatomical regions with a smaller region of interest, leveraging its enhanced learning capacity to improve segmentation accuracy.

Paper Structure

This paper contains 13 sections, 1 equation, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Learning Models for Medical Image Segmentation
  4. Femur Segmentation
  5. Materials and Methods
  6. Image Dataset and Pre-processing
  7. Deep Learning Models for Comparison
  8. Experimental Setup
  9. Training Setup
  10. Ensemble Approach
  11. Performance Metrics
  12. Experimental Results and Analysis
  13. Conclusion

Figures (2)

  • Figure 1: The figure presents segmentation results, with column (a) showing predictions from the SegmentAnyBone model and column (b) displaying results from the U-Net model. Predicted segmentations are marked in blue, while red boundaries denote ground-truth annotations. For SegmentAnyBone, the DSCs are 0.82, 0.92, and 0.56 from top to bottom. For U-Net, the corresponding DSCs are 0.69, 0.97, and 0.82.
  • Figure 2: Deep Learning Models Comparison for Different Parts.