Dr. McMillan, Principal Investigator of the MIMRTL group, in collaboration with Michael Fei, currently a medical student at Creighton University, have published a new paper in the Journal of Imaging Informatics in Medicine (JIIM) titled “Technical Note: Neural Network Architectures for Self-Supervised Body Part Regression Models with Automated Localized Segmentation Application.” This work presents advancements in self-supervised deep learning for medical imaging, particularly in the context of body part identification and localized segmentation. Additionally, the work demonstrates that organ segmentation algorithms trained on localized body regions perform better than a model trained on the entire body.
Addressing the Challenges in Whole-Body Medical Imaging
The increasing use of whole-body imaging modalities like CT, PET/CT, MRI, and PET/MRI has generated vast datasets. While these datasets offer comprehensive information, many clinical applications require focus on specific body parts or organs. Traditionally, deep learning models that perform such tasks rely on supervised learning with labeled data—a process that is both time-consuming and labor-intensive.
The Limitations:
- Manual Labeling: Requires significant effort and time.
- Computational Resources: Segmentation algorithms for whole-body scans can be resource-intensive.
- Scalability: Difficulty in scaling models to handle new regions of interest without retraining.
Introducing a Self-Supervised Solution
The paper extends self-supervised deep learning techniques that eliminate the need for manual labeling. By leveraging inherent data properties, the models learn to identify and localize anatomical structures efficiently.
Key Contributions:
- Comparison of Neural Network Architectures: Evaluated VGG, ResNet, DenseNet, ConvNeXt, and EfficientNet models for the body part regression (BPR) task.
- Automated Localization: Developed a pipeline using BPR slice scores to aid in anatomically localized segmentation.
- Improved Performance: Demonstrated that modern architectures like EfficientNet significantly outperform traditional models in the BPR task.
Methodology Overview
Dataset Used: Pediatric-CT-SEG dataset from The Cancer Imaging Archive, consisting of 359 CT scans with expert-labeled organ contours.
Body Part Regression (BPR):
- Objective: Predict a continuous slice score representing the relative position within the body.
- Approach: Models were trained to output slice scores without any manual labels, relying on self-supervised learning.
- Architectures Compared: VGG16 (baseline), ResNet50, DenseNet169, ConvNeXt_small, and EfficientNet_v2_s.
Loss Functions:
- Order Loss: Ensures that slice scores increase monotonically from inferior to superior slices.
- Distance Loss: Promotes linear spacing between slice scores for consistent localization.
Training Enhancements:
- Data Augmentation: Applied random transformations to increase variability and generalization.
- Random Slice Sampling: Selected random equidistant slices during training to account for varying slice thicknesses.
Significant Findings
- EfficientNet Outperforms Baseline: Achieved an overall mean absolute error (MAE) of 3.18, compared to 6.29 with the VGG baseline model—a 50.55% decrease.
- Localization Accuracy: EfficientNet localized organ landmarks within approximately 4% of the CT scan.
- Improved Segmentation Performance: Localized segmentation models using BPR slice scores significantly outperformed baseline models in 16 out of 20 organs, achieving a Dice similarity coefficient (DSC) of 0.88.
Implications:
- Resource Efficiency: Reduces computational requirements by focusing on localized regions, making it feasible to process large datasets.
- Scalability: The BPR approach can easily adapt to new anatomical regions without retraining extensive models.
- Enhanced Generalization: Models showed strong performance even on external datasets not seen during training.
- Performance of Localized Segmentation Models: Segmentation models trained on local body regions performed better than models trained on the entire body at once.
Practical Applications
- Automated Quality Assurance: Quickly verify the presence of correct body areas before intensive processing.
- Efficient Data Routing: Direct specific subregions to appropriate downstream algorithms for tasks like detailed segmentation or trauma detection.
- Improved Segmentation Models: Utilize BPR slice scores to develop focused segmentation models that require less memory and computational power.
Conclusion
The integration of modern neural network architectures and self-supervised learning presents a promising path for medical imaging. The ability to accurately and efficiently localize anatomical structures without manual labels holds significant potential for various clinical applications.
For access to the models visit the GitHub repository.