Enhancing Stroke Lesion Classification with Conditional GAN: Leveraging Unsupervised Clustering and Class-Conditioned Synthetic Data

This research study proposes an innovative approach to tackle the challenges of limited data availability and class imbalance in stroke lesion classification. We adopt a two-step methodology that combines unsupervised clustering with the Expectation-Maximization (EM) algorithm to obtain class labels and a Conditional Generative Adversarial Network (cGAN) to generate synthetic stroke lesion data. The EM algorithm aids in clustering the stroke lesion dataset into informative classes without the need for manual annotations, providing valuable class labels for subsequent data synthesis. Further, this study introduces a cGAN, incorporating these class labels, to generate diverse and realistic synthetic stroke lesion samples. The augmented dataset, comprising both real and synthetic data, is utilized to train a deep learning classifier for stroke lesion classification. The main objective of this research work is to enhance the classifier's performance by leveraging the benefits of both unsupervised clustering and cGAN-based data synthesis. Preliminary experiments demonstrate promising results, suggesting that the proposed methodology has the potential to improve stroke lesion classification accuracy and generalization. The proposed approach contributes to overcome the data scarcity challenges, making efforts towards achieving more reliable and effective stroke diagnosis and patient care.