Hallucinating for diagnosing: one-shot medical image classification leveraging score-based generative models

Deep learning models in data-scarce domains, such as medical imaging, often suffer from poor performance due to the challenges of acquiring large amounts of labeled data. Few-shot learning offers a promising solution to this problem. This work proposes a novel framework to jointly train a score-based generative model for high-quality sample hallucination and a meta-learning framework for one-shot classification. We evaluate our approach on MRI scans of prostate cancer, aiming to classify tumors based on severity. Our preliminary experiments demonstrate promising results, indicating the efficacy of our proposed method in improving classification performance. Future work will involve further analysis using a diverse set of score models and prototypical meta-learning techniques, as well as evaluation of the effectiveness of our framework in other medical imaging tasks.