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.

Authors

Abstract

To create a labeled dataset and evaluate a convolutional neural network (CNN) for segmentation of fluorine-18 sodium fluoride PET scans of the equine metacarpo- and metatarsophalangeal joint (fetlock), targeting the third metacarpal bone, proximal phalanx, proximal sesamoid bone(s), and soft tissue. PET and corresponding CT scans were retrospectively selected from June 2024 through November 2025 using convenience sampling and coregistered to a common reference frame. The coregistered PET-CT scans were manually labeled; the labels were then transferred to PET-only images and used to train a cascade of CNNs with and without data augmentation. Segmentation accuracy was quantified using the Dice coefficient and compared to atlas-based segmentation of the PET-CT scans. A total of 84 scans were collected from 2 facilities. For the test set (n = 8), the minimum, mean, and maximum Dice coefficients computed for all anatomical areas together were 0.85, 0.88, and 0.92, respectively, exceeding the values for atlas-based segmentation of 0.69, 0.80, and 0.87. Analysis of the Dice coefficient by area showed the best CNN segmentation for the third metacarpal bone (mean Dice coefficient (Dmean) = 0.92) and the worst for the medial proximal sesamoid bone (Dmean = 0.82); corresponding atlas-based values were 0.86 and 0.77. CNN-based segmentation of fluorine-18 sodium fluoride PET images of the fetlock outperforms an atlas-based method in both speed and accuracy measured by Dice coefficient. This advancement enables new, previously inaccessible strategies for PET image quantification. Accurate and robust segmentation of PET fetlocks enables more accurate analysis and novel insights into lesion characterization.

Tags

Topics