Abstract
This study introduces MedVis Suite, a framework developed to address key challenges in medical image analysis using MRI scans. MedVis Suite integrates advanced machine learning techniques, including a U-Net-based segmentation model optimized for bone segmentation, and 3D reconstruction capabilities. An in-depth evaluation of a U-Net-based model for bone segmentation is performed across anatomical planes, optimizing both loss functions and image scales. The axial view showed the highest performance with a Dice score of 0.91 using the baseline model, while the combination of Dice loss and boundary loss produced the best results. MedVis Suite offers significant potential to enhance medical image analysis, improve segmentation accuracy, and provide more comprehensive visualizations for clinical use. Future research will focus on validating MedVis Suite across diverse datasets and clinical applications, with the integration of image preprocessing techniques and fine-tuning strategies to further enhance the U-Net-based segmentation model.
Introduction
MRI is the gold standard for soft tissue visualization, yet it presents significant challenges for bone segmentation due to low signal intensity and poor contrast against surrounding tissues. Current manual segmentation workflows in tools like 3D Slicer are labor-intensive, prone to inter-observer variability, and often require high-end hardware. To address these limitations, we present MedVis Suite, an open-source framework designed to automate the "scan-to-model" pipeline for musculoskeletal analysis.
Methodology
The MedVis Suite provides a modular workflow encompassing DICOM exploration, automated segmentation, and 3D reconstruction. The core segmentation engine utilizes a U-Net architecture with a ResNet50 backbone, chosen for its optimal balance between accuracy and computational efficiency. The study systematically evaluated three critical variables:
- Anatomical Planes: Comparing performance across Axial, Coronal, and Sagittal views.
- Loss Functions: Testing eight configurations, including compound losses like Dice + Boundary (BD) loss.
- Image Scaling: Identifying the optimal input resolution to prevent overfitting.
Dataset and Experimental Setup
The study utilized ten distinct MRI datasets (TD01_S1_MRI through TD10_S1_MRI). Data was converted from Nearly Raw Raster Data (NRRD) and STL formats into PNG images for training. Notably, the Axial view provided the largest training set (e.g., 585 images for TD01) compared to the significantly smaller Coronal sets (166 images), creating a data distribution that influenced model performance across planes.
Results and Discussion
View Bias and Segmentation Performance
The model demonstrated a clear "view bias," with the Axial plane achieving a peak Dice score of 0.91, significantly outperforming Sagittal (0.73) and Coronal (0.69) views. This disparity is attributed to both the higher volume of axial training data and the more distinct anatomical boundaries visible in that plane.
Optimization Insights
- Compound Loss: The best results were achieved by combining Dice loss (region-based) with Boundary (BD) loss. This combination effectively penalizes the model for missing the general bone area while refining edge precision.
- The Augmentation Paradox: Contrary to standard computer vision practices, geometric augmentations like flips and rotations decreased performance, likely by disrupting the specific spatial relationships essential for anatomical recognition.
- Scaling Efficiency: Scaling analysis identified a "sweet spot" at an input scale of 0.6–0.7. Beyond this threshold, the model began learning noise, leading to overfitting rather than improved accuracy.
Conclusion
MedVis Suite offers a robust, user-friendly framework for automated bone segmentation, achieving high accuracy in axial views. Future work will focus on expanding dataset diversity to include various clinical pathologies and integrating advanced preprocessing techniques to improve performance in non-axial planes.
Summary updated