Analýza obrazu biomedicínských dat pomocí umělé inteligence

Abstract

This thesis develops practical biomedical image‑analysis pipelines that address three problems encountered in biomedical imaging: molecule extraction from optical‑mapping scans, instance segmentation of chronic lymphocytic leukaemia cells in bright‑field microscopy, and rib suppression in chest radiographs. For optical mapping, a preprocessing chain combining FOV alignment, U‑Net segmentation and watershed instance extraction produces BNX‑compatible molecule crops. For marker extraction two methods were developed, one is based on intensity of light and the other considers the Signal-to-Noise around the marker. For cell segmentation, a systematic comparison of modern encoder–decoder backbones identified U‑Net++ with a ResNeSt‑269 backbone as the best fit for the bright‑field data. For bone suppression, Autoencoders, U‑Net and GAN models with perceptual and gradient‑aware losses were trained to remove rib structures while preserving soft‑tissue detail.