Metody hlubokého učení pro automatické rozpoznání a klasifikaci lymfatických uzlin u metastáz karcinomu plic

Abstract

This diploma thesis focuses on the analysis of CT imaging data for the detection of metastatic involvement of lymph nodes in patients with lung cancer. The main objective was to design and evaluate an approach that enables patient classification based on imaging data without the need for invasive histological examination. First, different segmentation approaches were compared, including 2D and 3D convolutional neural networks. The 3D model based on the nnU-Net framework achieved significantly better results due to its ability to capture spatial context. The segmentation step was followed by patient classification, where multiple strategies were evaluated, including classical machine learning methods, ROI-based classification, and a model utilizing the entire CT volume combined with segmentation masks. The results show that classification at the patient level significantly outperforms approaches based on isolated ROI regions. The best performance was achieved using a combination of CT data, segmentation masks, and volumetric features. Furthermore, it was demonstrated that the use of automatically generated masks does not degrade classification performance, supporting the feasibility of a fully automated pipeline.