Abstract
Healthcare data are an essential resource in Machine Learning (ML) and Artificial Intelligence (AI) to improve clinical practice, empower patients, and enhance drug development with the aim to discover new medical knowledge. In particular, biomedical imaging analysis plays an important role in the healthcare context, producing a huge amount of data that can be used to study complex diseases and their evolution in a deeper way or to predict their onsets. In this work, we consider an approach based on Denoising Diffusion Probabilistic Models (DDPM), which is a type of generative model that uses a parameterized Markov chain and variational inference to generate synthetic samples that match real data. In particular, we execute a study by training on Malaria images and generating high-quality synthetic samples in order (i) to test the performance of the DDPMs, (ii) to estimate the association between original and synthetic data, and (iii) to understand how natural and human-made environmental factors impact Malaria disease. Finally, we use a well-defined convolutional neural network for classification tasks to assess the DDPM's goodness in generating synthetic images.
Introduction
In the field of medical AI, the availability of high-quality, annotated data is a primary bottleneck. Training robust deep learning models requires massive datasets, yet researchers frequently encounter significant obstacles:
- Privacy: Sensitive patient data is protected by strict legal frameworks like HIPAA or GDPR.
- Scarcity: Rare conditions often lack sufficient public samples for effective model training.
- Expertise: Annotating medical images, such as identifying malaria parasites in blood smears, requires specialized and time-consuming human effort.
This research investigates whether generative AI—specifically Denoising Diffusion Probabilistic Models (DDPMs)—can synthesize realistic medical data to overcome these challenges without compromising patient privacy.
Methodology
The study employs Denoising Diffusion Probabilistic Models (DDPMs). Unlike Generative Adversarial Networks (GANs), which can be unstable during training, DDPMs utilize a two-step process:
- Forward Process: Gradually adding Gaussian noise to a real image until it becomes pure static.
- Reverse Process: Training a neural network to iteratively "undo" this noise to recover a clean image.
Once trained, the model generates new synthetic images by starting from random noise and applying the learned reverse process. To evaluate the quality of these synthetic outputs, the researchers utilized YOLOv8, a state-of-the-art object detection and classification model, as a "blind judge" to classify the generated images.
Data and Experimental Setup
The study utilized a comprehensive Malaria dataset from the National Library of Medicine (NLM), containing 27,558 cell images balanced between "parasitized" and "uninfected" classes.
This dataset served as the ground truth for training the DDPM and the initial YOLOv8 classifier. Training was conducted over 100 epochs on an NVIDIA RTX 4090 GPU, demonstrating the feasibility of these models on consumer-grade high-end hardware.
Results and Statistical Validation
The DDPMs demonstrated high stability and reliability in generating complex medical textures. Key findings include:
- Classification Accuracy: The YOLOv8 model achieved a 97% accuracy rate on the malaria dataset, validating the realism of the synthetic samples.
- Statistical Fidelity: Chi-square and Kullback-Leibler (KL) divergence tests confirmed that the statistical distribution of the synthetic data closely matched the original dataset.
- Diversity: Synthetic images were found to be significantly different from the training set, indicating that the model learned underlying disease features rather than simply memorizing training examples.
Conclusion and Future Work
This work proves that DDPMs can generate high-fidelity synthetic medical images suitable for training diagnostic classifiers. Future research will focus on:
- Interpretability: Utilizing SHAP (SHapley Additive exPlanations) to identify the specific features (e.g., parasite morphology) the AI uses for classification.
- Environmental Factors: Exploring how geographic and environmental variables affect malaria presentation in images.
- Temporal Generation: Extending the approach to generate synthetic medical videos for surgical simulations and dynamic diagnostics.
Summary updated