Exploring artificial intelligence for differentiating early syphilis from other skin lesions: a pilot study.

BACKGROUND

Early diagnosis of syphilis is vital for its effective control. This study aimed to develop an Artificial Intelligence (AI) diagnostic model based on radiomics technology to distinguish early syphilis from other clinical skin lesions.

METHODS

The study collected 260 images of skin lesions caused by various skin infections, including 115 syphilis and 145 other infection types. 80% of the dataset was used for model development with 5-fold cross-validation, and the remaining 20% was used as a hold-out test set. The exact lesion region was manually segmented as Region of Interest (ROI) in each image with the help of two experts. 102 radiomics features were extracted from each ROI and fed into 11 different classifiers after deleting the redundant features using the Pearson correlation coefficient. Different image filters like Wavelet were investigated to improve the model performance. The area under the ROC curve (AUC) was used for evaluation, and Shapley Additive exPlanations (SHAP) for model interpretation.

RESULTS

Among the 11 classifiers, the Gradient Boosted Decision Trees (GBDT) with the wavelet filter applied on the images demonstrated the best performance, offering the stratified 5-fold cross-validation AUC of 0.832 ± 0.042 and accuracy of 0.735 ± 0.043. On the hold-out test dataset, the model shows an AUC and accuracy of 0.792 and 0.750, respectively. The SHAP analysis shows that the shape 2D sphericity was the most predictive radiomics feature for distinguishing early syphilis from other skin infections.

CONCLUSION

The proposed AI diagnostic model, built based on radiomics features and machine learning classifiers, achieved an accuracy of 75.0%, and demonstrated potential in distinguishing early syphilis from other skin lesions.