Mpox Detection Advanced: Rapid Epidemic Response Through Synthetic Data

The development and training of computer vision models for identifying visual symptoms of new diseases are significantly hampered by the dependence on large datasets. This is especially problematic during health crises where acquiring relevant data is challenging and time-consuming. Conventional data collection methods are inadequate for rapidly emerging diseases, leading to delays in developing detection models and potentially widespread disease. This paper introduces SynthVision, a novel approach using diffusion models to generate synthetic medical images from a minimal set of initial guide images to address this issue. While Generative Adversarial Networks (GANs) have been explored for synthetic image creation in healthcare, they suffer from limitations in diversity, unstable training, and a requirement for large datasets. Diffusion probabilistic models offer a significant advancement, demonstrating the ability to produce high-quality and diverse medical images, even under data scarcity. Recent improvements in Denoising Diffusion Probabilistic Models (DDPMs) have further optimized the image generation process, making them ideal for maximizing output from minimal inputs in vision model development. This paper aims to showcase the effectiveness of diffusion models in synthesizing images for healthcare to address challenges of sparse data availability and improve diagnostic precision.

Aljohani & Alharbe (2022) utilized Deep Pix2Pix GAN to generate synthetic medical images, demonstrating the potential of GANs. However, GANs are criticized for limited diversity and unstable training. Khader et al. (2022) showed that diffusion models can produce high-quality medical imaging data and enhance the performance of segmentation models under data scarcity. Nichol and Dhariwal's work (2021) on DDPMs highlighted improvements in image generation efficiency and quality. Ceritli et al. (2023) explored the use of diffusion models in creating realistic synthetic Electronic Health Records (EHRs), highlighting their flexibility. Kazerouni et al. (2023) provided a comprehensive survey of diffusion models in medical imaging. These studies collectively support the use of diffusion models for generating synthetic medical images to overcome data limitations.

The SynthVision methodology uses a customized text-to-image diffusion model, fine-tuned using the DreamBooth technique, to generate images of Mpox lesions. The process began with the curation of eight distinct sets of 15 clinically validated images each, comprehensively covering lesions on various body parts and skin tones (Fitzpatrick scale). Images were meticulously annotated with detailed descriptive texts to enhance the training process. 200 synthetic images were generated for each set, with 100-150 selected per set (totaling 1000 images) after clinical review. Both SDXL and SDv2 models were utilized, with fine-tuning via Dreambooth. A Vision Transformer (ViT) model was adapted for image analysis, processing images as 384x384 pixel inputs. The architecture was enhanced with attention dropout and an extra dense layer. The training dataset consisted of 1000 synthetic Mpox images, 1000 real images of normal skin, and 1000 real images of other skin disorders. A validation set (150 images per category) and a test set (100 images per category) of real patient images were used for model evaluation. Hyperparameter tuning optimized the model's performance, utilizing a batch size of 32, image size of 384x384 pixels, a learning rate of 1e-4, and dynamic learning rate reduction. The Adam optimizer and early stopping were employed.

The generated synthetic data realistically depicted Mpox lesions on various body parts and skin tones. The classification model achieved exceptional performance on the test set. For Mpox, the model achieved 96% precision, 96% recall, and an F1-score of 0.96. For normal skin, the model achieved 97% precision, 98% recall, and an F1-score of 0.98. For other skin disorders, the model achieved 97% precision, 96% recall, and an F1-score of 0.96. Overall, the model achieved 97% accuracy across all three categories (Mpox, Normal, Other). The confusion matrix and classification report confirm these results, demonstrating high true positive rates and low false positive rates across all categories.

The high accuracy and performance metrics demonstrate the effectiveness of the SynthVision approach in generating high-quality synthetic data for training a robust and accurate Mpox detection model. This addresses the critical challenge of data scarcity in rapid epidemic response. The use of synthetic data allows for quick model development without the delays associated with traditional data collection methods. The results suggest that diffusion models can effectively generate realistic and diverse medical images suitable for training computer vision models for disease detection. The consistent performance across different skin types and lesion locations enhances the model's generalizability and applicability in diverse clinical settings. This approach holds significant potential for rapid development and deployment of AI-powered diagnostic tools for emerging infectious diseases.

This study demonstrates the feasibility and effectiveness of using synthetic data generated by diffusion models for rapid development of accurate medical diagnostic tools. The SynthVision methodology achieved high accuracy (97%) in detecting Mpox lesions, even with a dataset composed primarily of synthetic images. This approach offers a valuable solution for addressing data scarcity challenges in time-sensitive medical emergencies. Future research could explore the application of this methodology to other diseases and the integration of multimodal data (e.g., combining image data with clinical information).

The study relied on a relatively small set of initial guide images for generating synthetic data. The generalizability of the model may be limited by the diversity of the initial dataset and potential biases in the source images. Further validation is needed with a larger and more diverse real-world dataset to confirm the model's robustness and performance under different conditions. While the model demonstrated high accuracy, its performance in real-world settings should be extensively evaluated before clinical implementation.