Recognition of Gouty Arthritis Using a Deep Learning Radiomics Model with Ultrasound Images: A Multicenter Study

Objective To develop and validate a hybrid deep learning-radiomics (DLR) model for characterizing sonographic features of gouty arthritis (GA) and compare its performance with radiomics (Rad) and deep learning (DL) models. Methods We analyzed ultrasound images from 714 patients with joint pain (training: 422; internal validation: 182; external validation: 110). Using Pyradiomics and ResNet50, we extracted 107 radiomic and 2048 deep learning features, respectively. Features were fused and optimized via LASSO regression to build the DLR model. Performance was assessed using AUC, DCA, and calibration curves. The model's GA detection capability was compared between junior and senior radiologists with/without DLR assistance. Results The DLR model achieved AUCs of 0.952 (training), 0.894 (internal validation), and 0.725 (external validation), significantly outperforming Rad models (P < 0.05). DLR assistance improved detection accuracy from 82.97–95.05% (P = 0.000) for junior radiologists and from 91.76–96.70% (P = 0.016) for senior radiologists, narrowing their performance gap (P = 0.375). Conclusion The DLR model shows promising potential for characterizing sonographic manifestations of gouty arthritis, improving recognition accuracy of GA-associated ultrasound features, especially for less-experienced radiologists.