Radiomics analysis for the early diagnosis of common sexually transmitted infections and skin lesions

Abstract

Early identification of sexually transmitted infection (STI) symptoms can prevent subsequent complications and improve STI control. We analysed 597 images from STIAtlas and categorised the images into four typical STIs and two skin lesions by the anatomical sites of infections. We first applied nine image filters and 11 machine-learning image classifiers to the images. We then extracted radiomics features from the filtered images and trained them with 99 models that combined image filters and classifiers. Model performance was evaluated by area under curve (AUC) and permutation importance. When the information of infection sites was unspecified, a combined Gradient-Boosted Decision Trees (GBDT) classifier and Laplacian of Gaussian (LoG) filter model achieved the best overall performance with an average AUC of 0.681 (95% CI 0.628-0.734). This model predicted best for lichen sclerosus (AUC = 0.768, 0.740-0.796). The incorporation of infection site information led to a substantial improvement in the model's performance, with 22.3% improvement for anal infections (AUC = 0.833, 0.687-0.979) and 3.8% for skin infections (AUC = 0.707, 0.608-0.806). Lesion texture and statistical radiomics features were the most predictive for STIs. Combining machine learning and radiomics techniques is an effective method to categorise skin lesions associated with STIs clinically.

Fig 1. Confusion matrices for the models predicted results on the testing dataset on (a) unspecific sites; (b) genitals; (c) anus; (d) other skin.

Fig 2. Models’ AUC (×10²) with nine images filter and eleven classifiers on (a) unspecific sites; (b) genitals; (c) anus; (d) other skin.

*ABC: Adaptive Boosting Classifier, DTC: Decision Tree Classifier, GBDT: Gradient Boosted Decision Trees, GNB: Gaussian Naive Bayes, GPC: Gaussian Process Classifier, KNN: K-nearest Neighbors, LR: Logistic Regression, MLPC: Multi-layer Perceptron Classifier, RC: RidgeClassifier, RFC: Random Forest Classifier, SVM: Support Vector Machine.

Fig 3. Comparison of the AUC of the best infection prediction model for infections on unspecific sites, genitals, anus and other skin.

Fig 4. The top rank 10 radiomics features in the best models using a permutation importance analysis.