Diffusion model uniform manifold filtering for classification of small datasets with underrepresented classes: Application to chromosomal aberration microscopy detection
Résumé
A frequent problem in biomedical machine learning is the issue of imbalanced classes, especially when datasets are small, which restricts the performance of deep learning methods. To address this issue, generative models are often used to generate additional synthetic data. Specifically, image-to-image models can transform input images to match the characteristics of target images. However, training such models on small datasets can affect the quality of synthetic samples. We propose a new method to filter generative outputs of an image-to-image Brownian Bridge Diffusion Models (BBDMs) using Uniform Manifold Approximation and Projection (UMAP) dimension reduction of a real data classifier's feature space. We apply this methodology to filter synthetic chromosomal aberrations generated from the blue DAPI-colored channels in the context of cytogenetic Fluorescence In Situ Hybridization (FISH) microscopy. Our method shows that such filtered synthetic data significantly enhance classification performance compared to the state of the art CycleGAN and could potentially be applied to a variety of other generative models.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
licence |