Artificial Intelligence Techniques and Developments for Infrared Imaging Based Bone Fracture Screening

This study developed and evaluated infrared thermal (IRT) imaging, coupled with artificial intelligence, as a technology for paediatric wrist-injury assessment. It addressed two persistent issues in emergency care: avoidable radiographs for soft-tissue injuries and unnecessary radiation exposure and associated cost. It was established that fractures and sprains which are clinically overlapping at presentation have distinct physiological signatures that manifest as significant temperature differences at the skin surface, which IRT imaging can detect reliably under a controlled acquisition. By developing and testing end-to-end pipelines, this study demonstrated the practical feasibility of IRT imaging-based screening as an adjunct to the current X-ray pathway. A review of conventional imaging (radiography, MRI, ultrasound) was undertaken and then IRT imaging positioned as a technology that can complement the existing diagnostics while reducing unnecessary radiation. Prior studies and early AI applications provided evidence that thermal features support fracture identification that was the focus of the study. To acquire the IRT images, a rigorous data-collection protocol was designed within Sheffield Children’s Hospital, ensuring ethical compliance, recording environmental control, and reproducible camera geometry. Forty children comprising 24 males and 16 females, mean age 10.50 years (standard deviation 2.63 years), 19 with a wrist fracture and 21 with a sprain were included for study with their injury type diagnosis (fracture or sprain) confirmed by an X-ray radiograph. Their mean body temperature was 36.3 oC (standard deviation 0.43 oC) across all participants. Bilateral, 10-second IRT videos were recorded and stabilised; anatomically standardised regions of interest (ROIs) covering carpal bones and distal radius/ulna enabled subject-wise, contralateral comparison that mitigates inter-patient variability. Three complementary feature extraction techniques were used, which included image pixel intensity statistics, a grid-based “hot-spot” representation, and frequency-domain (fast Fourier transform, FFT) magnitude features. These were used as part of training representative models: multilayer perceptron (MLP), convolutional neural network (CNN), adaptive neurofuzzy inference system (ANFIS), with appropriate normalisation, cross-validation, and targeted augmentation. The study’s findings provided a pilot analysis relating temperature to time-since-injury (TSI) and showed no early correlation for fractures but a subsequent moderate relationship, consistent with evolving inflammatory perfusion; sprains exhibited no significant temperature-TSI correlation. Secondly, an MLP trained on grid-based statistics achieved sensitivity 84.2%, specificity 71.4%, and accuracy 77.5%. This is a credible screening performance given the small cohort. Thirdly, a CNN processing FFT-transformed wrist ROIs, supported by rotation/translation/shear augmentation, provided an area under receiver operating characteristic (ROC) curve of 0.82 (accuracy 76%, sensitivity 88%, specificity 68%) and finally, a complementary ANFIS framework, trained on a compact triad of discriminative statistics (standard deviation, interquartile range, kurtosis), achieved effective separation of fractures and sprains while preserving interpretability via Takagi-Sugeno rules and membership functions. The study revealed that K-means-initialised Gaussian memberships offered the best balance of convergence speed and generalisation, outperforming fuzzy-C-means initialisations. Random initialisations of the ANFIS framework provided the best inference of the artificial intelligence models compared. In conclusion, the carefully standardised IRT, paired with thoughtfully chosen features and models, distinguished paediatric wrist fractures from sprains with clinically relevant performance, biological plausibility, and a credible path to explainable deployment. The evidence supported IRT imaging-based screening as a viable component in a modern fracture-care pathway, promising earlier decision-making, fewer unnecessary radiographs, and improved experience for paediatric care.