Deep learning cascade networks for segmentation of fluorine-18 sodium fluoride positron emission tomography scans of equine metacarpo- and metatarsophalangeal joints outperform atlas-based method.

Overview

• This study developed and evaluated a deep learning method using convolutional neural networks (CNNs) to segment fluorine-18 sodium fluoride PET scans of horse fetlock joints.
• The CNN-based segmentation outperformed traditional atlas-based methods in accuracy and speed, improving the analysis of equine joint PET images.

Background and Objective

• Fluorine-18 sodium fluoride PET imaging is used to assess bone and joint health in horses, specifically the equine metacarpo- and metatarsophalangeal (fetlock) joints.
• Accurate segmentation of key anatomical structures in PET images—such as the third metacarpal bone, proximal phalanx, proximal sesamoid bones, and surrounding soft tissue—is critical for lesion detection and analysis.
• Traditional atlas-based segmentation methods, which rely on pre-defined anatomical templates, can be limited in accuracy and computational speed.
• The study’s goal was to create a labeled dataset and assess whether a CNN cascade network could deliver superior segmentation performance compared to atlas-based approaches.

Data and Methods

• Researchers retrospectively collected 84 PET-CT scans from two veterinary facilities dated between June 2024 and November 2025 using convenience sampling.
• The PET and CT scans were co-registered to align them in the same spatial reference frame, facilitating accurate labeling.
• Manual labeling was performed on PET-CT coregistered images, identifying the relevant bones and soft tissues.
• These labels were then transferred to PET-only images, which are typically harder to segment due to lower anatomical detail compared to CT.
• A cascade of convolutional neural networks—a series of CNN models applied sequentially—was trained to perform segmentation, both with and without data augmentation techniques.
• Segmentation accuracy was measured using the Dice coefficient, a standard metric quantifying overlap between predicted segmentations and ground truth labels (values range from 0 to 1, with 1 indicating perfect overlap).
• Performance of the CNN approach was compared against an atlas-based segmentation method applied to the same PET-CT image set.

Results

• For the independent test set of 8 scans, the CNN-based segmentation achieved mean Dice coefficients of 0.88 across all anatomical regions, with a minimum of 0.85 and maximum of 0.92.
• The atlas-based method showed lower mean performance, with values around 0.80 on average (range 0.69 to 0.87), indicating the CNN approach substantially improved segmentation consistency and accuracy.
• Examining individual anatomical regions revealed:
  • The highest CNN accuracy was for the third metacarpal bone (mean Dice ≈ 0.92) versus atlas-based (≈ 0.86).
  • The lowest CNN accuracy was noted for the medial proximal sesamoid bone (mean Dice ≈ 0.82), still superior to atlas-based results (≈ 0.77).
• The CNN also provided segmentation at faster speeds than atlas-based methods, which often involve computationally intensive registration and fitting procedures.

Implications and Conclusion

• The cascade CNN approach provides accurate, robust, and expedited segmentation of fluorine-18 sodium fluoride PET images of equine fetlock joints.
• Improved segmentation enables more precise quantification of PET signals, which is essential for detecting and characterizing lesions in veterinary orthopedics.
• This method opens new possibilities for advanced image analysis and research into equine joint diseases by overcoming limitations of atlas-based segmentation.
• Overall, this work demonstrates that deep learning techniques can significantly advance the field of PET image analysis in veterinary medicine.