(P108) PAIRING SYNTHETIC IMAGE GENERATION WITH DISEASE CLASSIFICATION MODELS: SYNERGISTIC MACHINE-LEARNING APPROACHES PERMIT RAPID DEVELOPMENT OF DIGITAL DIAGNOSTIC TOOLS

Background:
Machine-learning models for disease classification have the potential to transform many domains of clinical medicine. Development of those models, however, depends on large image databases, which are not available for many diseases, labor- and time-intensive to develop, and difficult to access due to patient privacy protections. Pairing classification models with synthetic image generation may overcome the barriers to developing classification models, and permit their use in numerous contexts.

Methods:
Using 10 images of penises with human papilloma virus (HPV)-related disease, we trained a denoising diffusion probabilistic model. Pairing that model with a text-to-image model with text descriptors for a range of HPV-related disease presentations, we generated 630 synthetic images. Those images were assessed by an expert clinician for plausibility and quality, resulting in 500 synthetic images, which were used to train a downstream Vision Transformer model. We then assessed the model’s performance on 70 clinical images of penises with HPV-related disease, 70 on diseases other than HPV (herpes simplex virus (HSV) (n=30), primary syphilis (n-20), penile cancer (n=10), penile eczema (n=5), and penile psoriasis (n=5)), and 70 penile images without disease, calculating recall (or sensitivity), precision, specificity, and F1-score, as well as the Area Under the Receiver Operating Characteristics Curve (AUC).

Results:
The model correctly classified 64 of 70 images of HPV-related disease, with a recall (or sensitivity) of 91.4% (94% CI 82.3% – 96.8%). The precision of the model for HPV-related disease was 95.5% (95% CI 87.5% - 99.1%), and the F1-score was 93.4%. The model AUC was 93.1% (95% CI 88.3% - 97.9%).

Conclusion:
We paired denoising diffusion probabilistic modeling with a text-to-image model to produce synthetic images based on minimal initial input, which we used to train a downstream disease classification model. That classification model demonstrated excellent performance for classifying images of penile pathology.