Despite increasing prevention and control measures, sexually transmitted infections (STIs) are increasing in many countries, including Australia. Early detection of STIs is crucial to prevent severe clinical consequences and further spread throughout the community. Our study aims to develop an artificial intelligence (AI)-assisted diagnostic tool to help the public and clinicians detect the lesions early.
Methods:
We used the existing STI lesion images recorded at Melbourne Sexual Health Centre (MSHC). We included a total of ~4,000 (70% for training, 20% for validation and 10% for testing) lesion images caused by STIs (genital warts, herpes, syphilis, molluscum contagiosum, monkeypox, gonorrhoea) and non-STIs (healthy skin, pearly white penile papules, vaginal intraepithelial neoplasia (VIN), balanitis, lichen sclerosis and other dermatoses). We performed lesion alignment of images and used a residual convolutional neural network (ResNest-34) to learn coupled spatiotemporal features from aligned images. Then, we extracted the spatiotemporal features of the lesions to classify them into the respective STIs in comparison with non-STIs.
Results:
The best-performed models with 150 epochs, and a 3e-4 learning rate, achieved overall model accuracy values at 86.9% when it was evaluated with 20% of the validation dataset. Among STI lesions, the model achieved accuracy values at 94%, 91%, 87%, 88%, 79% and 76% for molluscum, syphilis rash, genital warts, monkeypox, genital herpes and syphilis ulcer respectively. Among non-STI lesions, accuracy was achieved at 98.3%, 94.7%, 90%, 85% and 81.6% for penile papules, dermatosis, healthy skin, VIN and lichen sclerosis respectively. The model performance was evaluated again with a 10% testing dataset, and it achieved similar accuracy at 86.8%.
Conclusion:
The AI-assisted diagnostic tool could be an effective measure to assist the public and clinicians in the early detection of STIs. As a future work, MSHC would collect more images to construct a more efficient model and make the tool available to the public.
