Modelos de Difusão Latente na Super-Resolução de Imagens Mamográficas

Abstract

In recent years, generative models have gained popularity for various tasks, including text and image generation. In this thesis, I will develop a generative model to improve image resolution of mammography simulations using super-resolution techniques based on latent diffusion with AutoencoderKL methods. Super-resolution task involves increa sing the resolution of an image, typically from a low-resolution image, often with loss of quality, aiming to generate a high-resolution version that enhances visual quality. It is particularly useful in fields like medical image processing, where precision is crucial for accurate diagnoses. In the context of mammography, super-resolution can help reveal subtle details that may go unnoticed in low-resolution images, thereby contributing to more precise anomaly detection and the development of more effective treatments. The data consists of two sets of unlabeled images, one in high resolution and the other in low resolution, both containing 625 grayscale images. The results were evaluated using FID, SSIM, and PSNR, achieving values of 0.00006, 0.80, and 26, respectively in the super-resolution task, which are satisfactory and comparable to the literature. However, due to the heterogeneity and simulated nature of the data, the AutoencoderKL model produced images with slightly degraded images. Further study is suggested to enhance the model’s effectiveness by exploring different training strategies, normalization tech niques, and alternative Autoencoders for latent representation generation. Additionally, this work contributes to the community by proposing a new pipeline for training and eva luation in the super-resolution task, with containerized code for scalable reproducibility across different operating systems and datasets.