TL;DR
This paper introduces a novel, iterative framework for customizing pre-trained deep learning models to automate the segmentation of abdominal fat (visceral and subcutaneous) from hybrid PET/MRI scans. This approach significantly reduces annotation time and provides highly accurate volumetric and metabolic measurements, crucial for understanding conditions like Obstructive Sleep Apnea (OSA).
What Problem Does This Paper Address?
Accurately measuring visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT) is vital for understanding the health risks associated with obesity, particularly in conditions like Obstructive Sleep Apnea (OSA). Traditional methods for segmenting these fat depots from medical images (like MRI and PET) are time-consuming and labor-intensive, requiring expert manual annotation. While deep learning (DL) offers automation potential, developing models tailored to specific research needs and imaging modalities (like hybrid PET/MRI) can be challenging, especially with limited labeled data. This paper addresses the need for an efficient, customizable framework to automate abdominal fat segmentation from PET/MRI, thereby accelerating research into obesity, OSA, and related cardiometabolic diseases.
Methods
The core of this work is a "continuous closed-loop" AI model training strategy. They leveraged a pre-trained UNet-ResNet50 model, initialized with weights from RadImageNet (RIN), a large medical imaging dataset. This model was then iteratively refined using a custom framework within the "Discovery Viewer" platform.
Here's the breakdown:
- Transfer Learning: Starting with a model pre-trained on a diverse medical imaging dataset (RadImageNet) provided a strong foundation, reducing the need for extensive training from scratch.
- Iterative Training (V1, V2, V3): The model was trained in stages:
- V1: Trained on an initial set of 59 manually annotated scans.
- V2: Refined using 157 scans, incorporating corrections from the V1 model's predictions.
- V3: Further refined with 328 scans, representing the final model.
- Closed-Loop Annotation: This is a key innovation. Expert reviewers corrected the AI's segmentations, and these corrected annotations were fed back into the training dataset for the next iteration. This "annotation-correction workflow" significantly reduced the overall manual annotation burden.
- Discovery Viewer Platform: This platform facilitated the integration of the AI model into the annotation workflow, allowing for real-time visualization, annotation, and refinement of AI predictions.
Datasets
The study utilized a dataset of hybrid 18F-FDG PET/MRI scans from two prospective sleep cohorts at the Mount Sinai Health System.
- Training Data: A total of 328 annotated PET/MRI scans were used for training the models (V1, V2, and V3). The iterative process started with 59 scans for V1, expanded to 157 for V2, and concluded with 328 for V3.
- Test Data: 10 independent test cases were used for evaluation. Each test case comprised 80-100 MR slices.
The paper does not introduce a new benchmark dataset but rather leverages existing patient data for model development and validation.
Key Insights and Ideas
- Customization is Key: The paper strongly advocates for customizable AI pipelines over universally generalizable models. Research projects often have unique requirements (e.g., specific slice ranges, exclusion criteria), making tailored models more practical and efficient.
- Iterative Refinement Accelerates Progress: The "continuous closed-loop" approach, where AI predictions are corrected and fed back into training, is a powerful way to rapidly improve model performance while drastically reducing manual annotation effort.
- Transfer Learning for Medical Imaging: Pre-training on large, diverse medical imaging datasets like RadImageNet provides a significant head start, enabling models to achieve good performance even with relatively smaller, specialized datasets.
- Mask vs. Contour Accuracy: While the AI models achieved excellent accuracy for segmenting the overall fat masks (crucial for volume and metabolic measures), the accuracy for delineating precise contours was lower. This highlights a trade-off and suggests areas for future improvement.
Key Results
The proposed framework demonstrated substantial improvements in both efficiency and accuracy:
- Time Efficiency:
- Manual segmentation: 121.8 minutes per scan.
- AI-assisted segmentation (V3): 31.8 minutes per scan.
-
Fully automated AI segmentation (V3): 1.2 minutes per scan. This represents a significant reduction in annotation time, with the AI-assisted approach being ~4x faster than manual.
-
Segmentation Accuracy (Dice Similarity Coefficient - DSC):
- Masks: Consistently high DSC values (0.98 for VAT and SAT regions) across all model iterations, indicating excellent overlap with manual annotations.
- Contours: Lower DSC values (averaging 0.43 for VAT and 0.54 for SAT contours), suggesting room for improvement in boundary delineation.
- Volumetric and Metabolic Agreement:
- Intraclass Correlation Coefficients (ICCs): Excellent agreement between AI-derived and manual measurements, with ICCs consistently exceeding 0.98 for the best model (V3) for both volumetric and metabolic measures.
- Bias: Minimal systematic bias was observed for VAT/SAT SUVmean and VAT/SAT ratio. A small, clinically negligible bias was noted for total VAT and SAT volumes, attributed to contour imprecisions.
Limitations and Future Work
- Contour Accuracy: The authors acknowledge that contour-level DSC values were lower than mask DSC values. While this didn't significantly impact volumetric or metabolic outcomes, it indicates an opportunity for further optimization in boundary delineation.
- Future Work for Contour Improvement: Strategies like incorporating edge-aware or boundary-refinement loss functions (e.g., Boundary IoU, Hausdorff distance) are suggested to enhance contour precision.
- Sample-Selection Bias: The iterative, expert-corrected approach might introduce sample-selection bias or overfitting to an annotator's style. However, they mitigated this by demonstrating high inter-reader reliability in prior work.
- Generalizability: The model's generalizability across diverse datasets and imaging conditions remains to be validated. The primary focus was on annotation efficiency for specific research needs.
- Future Platform Development: The Discovery Viewer platform is planned for a public, cloud-hosted release to support broader collaboration and community-driven model development.
Summary updated