A Two-Stage Generative Model for Unsupervised Brain Tumor Detection Using CycleGAN and Diffusion Models

Brain tumor detection from MRI images is a critical task in medical image analysis, yet it remains challenging due to limited labeled data, complex tumor structures, and high annotation costs. Traditional supervised learning methods depend heavily on large annotated datasets, which are often difficult to obtain in real clinical scenarios. To address these challenges, this research proposes a two-stage unsupervised generative learning framework for brain tumor detection and segmentation. In the first stage, a CycleGAN model is employed to generate synthetic abnormal MRI images [3] from healthy MRI images, creating pseudo-paired healthy–abnormal data without manual annotation. In the second stage, a conditional diffusion-based generative model is applied [4],[5] to reconstruct healthy images from abnormal inputs by learning the joint distribution between healthy and abnormal image pairs. The difference between the reconstructed healthy image and the original abnormal image is used to localize tumor regions. The proposed system is evaluated using confusion matrix–based performance metrics including precision, recall, F1-score, accuracy, and Dice similarity coefficient. Experimental results demonstrate that the proposed framework effectively detects and segments tumor regions, achieving reliable performance without requiring labele training data. This approach provides a robust, scalable, and annotation-free solution for medical image anomaly detection and brain tumor analysis.