Gürültü giderici difüzyon olasılık modelleri ile iç boyamayoluyla beyin fraksiyonel anizotropi haritalarındadenetimsiz anormallik tespiti

Anomaly detection constitutes a critical area of investigation across various research domains. Recent advancements in deep learning have facilitated the application of neural networks for anomaly detection. However, the high costs and time-intensive nature of data collection and annotation have driven the emergence of unsupervised learning techniques. Within the realm of unsupervised learning, generative models have gained prominence, enabling the generation of diverse data for training purposes. In the context of medical imaging, specifically brain MRIs, generative models have been utilized to detect anomalies in 2D slices of 3D volumes, typically employing structural modalities like T1-weighted or T2-weighted images. This study employs a generative Denoising Diffusion Probabilistic Model (DDPM ) to produce 3D partial volumes of brain Fractional Anisotropy (FA) maps by applying inpainting in inference, which are subsequently masked with inpainting masks to generate the full 3D volume. Anomalies are identified by assessing the disparity between the model's input and output. The performance of the generative model is evaluated using metrics such as Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), Maximum Absolute Difference, and Average Percentage Error. Anomaly detection accuracy is assessed using Intersection over Union (IoU) and F1 score. The proposed model achieved 0.05 IoU and 0.095 F1 score. The methodology also compared with prior work and it has been shown that while prior works achieved %50 DICE score, while inpainting in inference achieved %39.01 DICE score.