Objective assessment and virtual lesion simulation of psoriasis lesion severity based on deep learning regression and generative models

Astratt

Objective The Psoriasis Area and Severity Index (PASI) remains the gold standard for evaluating psoriasis severity; however, the assessment of its subscores—erythema, scaling, and induration—relies heavily on clinicians’ subjective judgment, leading to poor inter-rater consistency. This study aims to develop an objective evaluation model, Psoriasis-RegNet, based on deep learning regression, using publicly available dermatological image repositories, and to explore a conditional generative model, Psoriasis-GAN, for creating virtual lesion simulations of psoriasis lesions and demonstrating their potential applications in medical education. Methods A total of 1,250 clinical images of psoriasis were collected from the public DermNet dermatology image archive. Based on the accompanying clinical descriptions, a large pre-trained language-vision model was employed to automatically annotate each image with erythema ( E ), scaling ( S ), and induration ( I ) scores on a 0–4 scale, thereby constructing a multi-label regression dataset. The reliability of the pseudo-labels was evaluated against independent annotations from board-certified dermatologists, yielding ICC values of 0.86, 0.83, and 0.79 for erythema, scaling, and induration, respectively. (1) Psoriasis-RegNet: We developed a multi-task regression network using ResNet-50 as the backbone, pretrained on the ISIC 2019 skin lesion dataset. Three independent regression heads were attached to predict the E , S , and I scores, respectively, enabling simultaneous quantitative evaluation of lesion severity components. (2) Psoriasis-GAN: We further implemented a conditional generative adversarial network (cGAN) based on the StyleGAN2-ADA architecture. The composite vector E,S,I was used as the conditional input to guide the generation of high-resolution psoriasis lesion images, supporting controllable and clinically meaningful visual synthesis. Results The Psoriasis-RegNet model demonstrated excellent performance on the held-out internal test set. The mean absolute errors (MAE) for erythema, scaling, and induration predictions were 0.42, 0.45, and 0.48, respectively. The Pearson correlation coefficients (PCC) between model predictions and the held-out reference labels were 0.85, 0.82, and 0.80 for erythema, scaling, and induration, respectively (all P < 0.001). These prediction errors are within the range of inter-rater variability reported among dermatologists for PASI subscores (typical MAE: 0.5–0.9). The Psoriasis-GAN successfully generated high-fidelity (1,024 × 1,024) and conditionally controllable lesion images, achieving a Frechet Inception Distance (FID) of 28.5, and a Visual Turing Test discrimination accuracy of 51.3 percent, indicating near-photorealistic quality. Conclusion This study demonstrates that re-annotation of public dermatological image repositories provides a feasible approach for developing AI models in dermatology. The proposed Psoriasis-RegNet offers an objective and quantitative tool for psoriasis severity assessment, while Psoriasis-GAN shows great potential for medical education and AI data augmentation through the creation of virtual lesion simulations that enable interactive exploration of psoriasis severity phenotypes.