Diabetic retinopathy is an eye complication caused by a widespread disease named diabetes mellitus. The examination of retinal fundus images procured by retinography is the most commonly used method for diagnosing diabetic retinopathy. Strategies based on deep learning have shown promising results in detecting diabetic retinopathy, achieving performance similar to that of the human eye regarding image inspection. However, the performance of these strategies heavily depends on fine-tuning the algorithm hyper-parameters and big datasets. In this work, we propose training a Deep Learning model with Low-Rank Adaptation (LoRA) approach to classify three stages of Diabetic Retinopathy: i) no sign of diabetic retinopathy, ii) Non-proliferative diabetic retinopathy, and iii) proliferative diabetic retinopathy. We propose using a low-rank representation to reduce significantly the number of trainable parameters. The experiment shows that the LoRA approach for image classification of the three stages of diabetic retinopathy manages to obtain state-of-the-art results even with a small dataset.

Diabetic retinopathy is an eye complication of a widespread disease named diabetes mellitus. The most widely used method for diagnosing diabetic retinopathy is the analysis of retinal fundus images obtained by retinography. Deep Learning-based methods have shown promising results as a diagnostic tool for diabetic retinopathy, achieving, in some cases, performance close to the human inspection of images. However, the performance of these methods relies heavily on fine-tuning the algorithm hyperparameters and big data sets. In this work, we propose training a Deep Learning network with evolutionary algorithms to classify three stages of Diabetic Retinopathy: i) no sign of diabetic retinopathy, ii) Non-proliferative dia-betic retinopathy, and iii) proliferative diabetic retinopathy. We propose a neuroevolution methodology for selecting the most efficient Deep Learning model. The results of the neuroevolution methodology were improved by including Simulated Annealing strategies, Population Reinitialization, and ensembles. With high accuracy, sensitivity, specificity, and kappa index rates of 0.889, 0.889, 0.951, and 0.822, respectively, in the best case found, the experiments show that our neuroevolution methodology for selecting the Deep Learning model hyperparameters is a competitive alternative for training deep neural networks to classify three stages of diabetic retinopathy even with a small data set.

Image segmentation is one of the first steps in most image processing procedures. The segmentation aims to obtain a more meaningful or simplified image representation by grouping pixels with common characteristics, which allows regions or features of interest to be uniquely identified. The result of the segmentation has a significant impact on the subsequent steps. Segmentation is part of several superior applications such as artificial vision, medical, topographic, and astronomical image analysis. No single or universal segmentation process gets optimal performance for all image types. Hence, determining a function that fits specific image types or applications becomes a detailed, complex, and not trivial task requiring much time and effort. In this paper, we propose using Multi-Objective Evolutionary Algorithms (MOEAs) as a training tool that combines operations that represent the techniques and strategies commonly used for generating image segmentation. As a result, sequences of operations are suitable for specific applications or image types. The objective functions used to guide the evolutionary process are sensitivity maximization (TPR) and specificity maximization (TNR), the basic components of ROC analysis. Sensitivity and specificity are commonly used as classification metrics to evaluate the quality of a proposed segmentation compared to an ideal segmentation. We used sensitivity and specificity as objective functions rather than accuracy because, as stated in [1], the dependence on prevalence makes accuracy less effective than a simultaneous consideration of sensitivity and specificity. Experiments were conducted on multiple images that share common characteristics obtained from image databases, specifically: i) benign and malignant melanoma images, ii) ophthalmoscopic retinal images, and iii) binary cell form images, where the segmentation generated by the proposed algorithm was compared with ideal segmentation. The results are quite promising and show that using MOEAs to generate sequences of segmentation operations valid for specific applications is feasible.