<p>Deep learning (DL) has become one of the promising methods of automated diagnosis and grading of diabetic retinopathy (DR). Nevertheless, it is the rapid proliferation of proposed models that has created diverse evidence on their performance, clinical relevance and practical applicability. The aim of this paper is to address this challenge by providing a systematic review of the deep learning techniques for DR detection from 2020 to 2025. After following the PRISMA selection process, 98 studies were included. The literature review is divided into four key sections, including single-branch convolutional neural networks (CNNs), hybrid and ensemble models, multimodal methods, as well as transformer-based models. The results indicate that CNN-based models are still computationally inexpensive and clinically viable, and hence can be used in large-scale screening. They, however, are less sensitive to early-stage DR and also tend to have difficulties in extrapolating between datasets. Conversely, the hybrid, multimodal, and transformer-based models show an improved level of diagnostic performance with accuracy ranging from 95 to 98% depending on the task and dataset. These models, however, have relatively low external validation and will need more data and more computing. The literature also suffers from a number of methodological issues, such as data leakage, inconsistent validation strategies and exaggerated performance reporting, which cast doubt on the validity and practical relevance of most of the models. In general, this paper presents a critical review of existing deep learning-based DR detection systems and highlights the need to improve current approaches. A shift from performance-driven evaluation toward efficient, explainable, and clinically applicable models has been observed. The study also outlines future research directions, including the need for robust validation, multidimensional datasets, and scalable solutions for practical healthcare implementation.</p>