Automatic quantification of thermographic images of complex regional pain syndrome using radiomics and deep learning.

INTRODUCTION

Complex regional pain syndrome (CRPS) is a complication after trauma or surgery. Skin temperature asymmetry is one of the few symptoms that can be measured, for which thermography can be used. However, the current assessment of thermograms relies on manual interpretation and is subjective.

OBJECTIVE

The aim of this study was to develop an automatic model for quantitative assessment of CRPS based on thermography by using deep learning and radiomics.

METHODS

In this study, 178 thermograms of the extremities of 98 patients with CRPS and 837 thermograms from 56 healthy controls were included. A deep learning model was developed to segment the extremities. From each thermogram, for each extremity, 564 radiomics features were extracted. Based on these features, a classification model was developed using a combination of machine learning approaches and evaluated through a 20x random-split cross-validation. The performance of the classification model was compared with visual scoring of the thermograms by 3 clinicians.

RESULTS

The radiomics classification model statistically significantly outperformed all 3 clinicians (mean area under the curve of 0.93 compared with 0.82, 0.79, and 0.69 [P < 0.001]).

CONCLUSION

Our automatic classification model can distinguish thermograms of patients with CRPS from those of healthy controls with a performance that exceeds that of clinicians, thereby providing a basis for future studies evaluating its role in diagnosis and treatment monitoring.