Convolutional neural networks for classifying three-phase bone scintigraphy after total knee arthroplasty: A pilot study
Numan Mercan, Ahmet Yurteri, Mustafa Erol, Muhammed Tekin ErtekinObjectives
Total knee arthroplasty (TKA) is increasingly performed, with complications including pain, loosening, and periprosthetic joint infection. When the cause is unclear, three-phase bone scintigraphy (3PBS) is used to assess uptake patterns categorized as normal, aseptic, or septic involvement. We aimed to develop a CNN-based model that classifies phase-wise uptake patterns labeled by a nuclear medicine specialist; side-specific case-level categories were then assigned using a prespecified rule-based algorithm based on the phase outputs.
Materials and Methods
This retrospective single-center study included 102 patients (81 females, 21 males; mean age 67.0 ± 7.5 years, range 47–84) who underwent 3PBS at least 12 months after TKA (24 right, 25 left, 53 bilateral). Each scan included blood-flow, blood-pool, and delayed phase images labeled. Experiments using five CNN models (ResNet50, VGG11, VGG16, EfficientNet-B0, MobileNet) on phase- and laterality-specific inputs produced low accuracy (33–51%). Therefore, we developed customized CNNs for each phase–laterality combination. Predictions from these models were integrated using a decision algorithm for final side-specific case-level classification.
Results
The customized CNNs achieved 48/49 (98.0%) accuracy in the blood-flow phase, 47/48 (97.9%) in the blood-pool phase, and 50/52 (96.2%) in the delayed phase. Combining phase outputs yielded a side-specific case-level accuracy of 43/47 (91.5%) among test cases with complete three-phase data.
Conclusion
A phase-specific, laterality-aware CNN architecture showed high accuracy in classifying specialist-labeled 3PBS uptake patterns after TKA, with side-specific case-level categories assigned using a prespecified rule-based algorithm based on the phase outputs. These findings are promising but preliminary within this single-center pilot dataset.