Аңдатпа
This work is intended to study methods for pre-processing and analysis of fundus images for the detection of diabetic retinopathy. Diabetic retinopathy (DR) is a common eye disease in patients with diabetes, and its early diagnosis allows you to prevent vision loss. During the study, modern methods for processing and analyzing fundus images were used, including the EfficientNetB0 architecture based on Deep Learning. Image augmentation (rotation, scaling, cropping, contrast enhancement) and normalization methods were introduced for pre-processing. When using the EfficientNetB0 architecture, two approaches were tested: training the base layers and additional adaptation (fine-tuning) by opening the upper layers. The results were evaluated by metrics. The precision for the test set in the first method was 65%, and for the second method 75%. The accuracy of the validation set in the first method was 63%, and in the second method it reached 71%. The recall metric showed 60% for the test set in the first method, and 74% in the second method. In general, the fine-tuning method showed high performance. The use of these methods allows to improve the quality of image processing and classification for effective diagnosis of diabetic retinopathy. The novelty of the study is the analysis of various methods of using and adapting the highly efficient EfficientNetB0 architecture. The results obtained allow to improve the quality of automated systems in DR diagnostics and increase the energy efficiency of the model. The proposed methods have high potential for early detection of eye diseases.