Metody głębokiego uczenia w rozpoznawaniu chorób siatkówki : Od augmentacji danych do transformerów wizyjnych
The monograph covers the problem of automatic recognition of rare retinal diseases using advanced deep learning methods. The work constitutes a coherent, multi-faceted scientific study that integrates issues of medical data engineering, the design of neural network architectures, methods for improvi...
Saved in:
| Hovedforfatter: | |
|---|---|
| Format: | Online |
| Sprog: | engelsk |
| Udgivet: |
Lublin University of Technology Publishing House
2026
|
| Fag: | |
| Online adgang: | https://directory.doabooks.org/handle/20.500.12854/176032 |
| Tags: |
Ingen Tags, Vær først til at tagge denne postø!
|
| Summary: | The monograph covers the problem of automatic recognition of rare retinal diseases using advanced deep learning methods. The work constitutes a coherent, multi-faceted scientific study that integrates issues of medical data engineering, the design of neural network architectures, methods for improving learning efficiency under limited data availability, and the interpretability of artificial intelligence models in clinical context. The monograph fills a significant research gap related to the lack of effective, reliable and explainable tools supporting the diagnosis of rare ophthalmic diseases, such as retinitis pigmentosa and acquired vitelliform lesion. The primary objective of the monograph is to develop and comprehensively evaluate modern deep learning methods capable of accurate identification of rare retinal diseases based on ophthalmic images. The author seeks to enhance diagnostic performance under conditions of limited data availability, high phenotypic variability, and subtle pathological changes. The specific objectives include: (i) the development of advanced data augmentation strategies, including those based on generative adversarial networks; (ii) the design of original convolutional neural network architectures and vision transformers; (iii) the reduction of computational complexity while maintaining high classification performance; (iv) the integration of ensemble learning methods; and (v) ensuring the interpretability of model classification in clinical context. The monograph is based on a broad spectrum of computational and experimental methods. First, the author provides a detailed analysis of the characteristics of selected rare retinal diseases and describes the process of constructing and preparing dedicated datasets. A key methodological component is an extensive data augmentation module. In addition to classical geometric and photometric transformation techniques, the author proposes the use of Deep Convolutional Generative Adversarial Network (DCGAN) to generate synthetic retinal images, which significantly alleviates the problem of data scarcity and class imbalance. With regard to identification architectures, the monograph covers both an analysis of reference models (ResNet, DenseNet, Inception) and original solutions. Particular attention is paid to the Deep CNN–GRU model, which combines spatial feature extraction with contextual dependency modelling, the convolutional GRU U-Net for simultaneous segmentation and classification, and a dedicated Residual Attention Network that enables selective focus on key pathological regions. An important methodological contribution is the introduction of Kronecker convolution as a substitute for classical convolutional filters, allowing a reduction in the number of parameters and improved learning efficiency while dealing with limited datasets. The subsequent part of the study covers vision transformers, including the author’s Deep Residual Vision Transformer architecture, enhanced with a residual self-attention mechanism. The author analyses the image tokenisation process, single-head and multi-head attention mechanisms, and the impact of token positioning on classification quality. Ensemble methods (bagging, boosting, stacking, soft voting) are employed to further improve identification stability and accuracy. The study is complemented by explainable AI techniques, in particular Grad-CAM and SHAP, which enable visual and quantitative interpretation of model decisions. The conducted experiments demonstrate that the application of advanced data augmentation using DCGAN leads to a significant improvement in classification performance compared to classical augmentation methods. The proposed CNN–GRU, RAN and Deep Residual ViT architectures achieved high accuracy, sensitivity and F1-score, exceeding 95%, while maintaining good generalisation capability. Kronecker convolution reduced the computational complexity of the models without compromising performance and, in many cases, resulted in further improvements. The use of ensemble learning additionally increased classification stability, particularly under class imbalance conditions. Statistical analysis using McNemar’s test confirmed the significance of the observed improvements relative to baseline models. Explainability analyses showed that the models focus their attention on retinal regions consistent with established clinical knowledge, thereby increasing trust in the proposed solutions and their potential applicability in clinical practice. The monograph demonstrates that the combination of advanced data augmentation techniques, modern deep learning architectures, attention mechanisms, computational complexity reduction, and explainable methods constitutes an effective strategy for the recognition of rare retinal diseases. The proposed solutions significantly outperform classical approaches while offering transparency and clear potential for clinical deployment. The study makes an important contribution to the development of medical artificial intelligence and outlines directions for future research, including adaptive generative models, hierarchical vision transformers, and integration with multimodal clinical data. |
|---|