In this study, the impact of synthetic noise on the classification accuracy of HEp-2 cell images was analyzed using a deep convolutional neural network based on the YOLO architecture. The nature and origin of three prominent types of noise: photon, Gaussian, and impulse – were investigated in the context of fluorescence microscopy. The robustness of a modern neural classifier to each noise type was examined through quantitative performance evaluation across six staining patterns of antinuclear antibodies. The influence of noise on morphological feature representation and class-specific recognition was studied using standard classification metrics. A methodology for generating noise perturbations in diagnostic images was proposed and implemented. The architecture and training configuration of the selected model were described and substantiated in detail. Classification experiments under various image quality conditions were conducted and analyzed. The approach formulated in this work characterizes a practical framework for assessing the reliability of computer-aided diagnostic tools under realistic imaging constraints.