Automatizovaná klasifikace sítnicových snímků novorozenců na základě přítomnosti hemoragií

Abstract

This bachelor’s thesis focuses on the development and implementation of a classification algorithm for the automated detection of retinal hemorrhages in newborns using deep learning methods. The main objective of the thesis was to design a convolutional neural network model capable of classifying retinal hemorrhages in ophthalmic images. The ResNet-18 architecture was selected as part of the methodology, with the principle of transfer learning being applied. Given the significant imbalance in the clinical dataset used, techniques of geometric data augmentation and loss function weighting were employed during the training process. The developed model demonstrates high effectiveness with an overall accuracy of 96.11 %, sensitivity of 92.31 %, and specificity of 96.50 %. A major contribution of this work is also the development of a diagnostic classification application in the MATLAB App Designer environment, which includes visualization of decision-making processes using the Grad-CAM method and enables the generation of PDF reports.