Abstract
Building rehabilitation is a reality; all rehabilitation work phases must be efficient and sustainable and promote healthy living places. Current procedures for assessing construction conditions are time-consuming, labour-intensive, and costly. They can threaten engineers' health and safety, especially when inspecting hard-to-reach and high-altitude sites. At the initial stage, a survey of the condition of the building is conducted, which later implies the preparation of a report on the existing pathologies, intervention solutions and associated costs. This procedure involves an inspection of the site (through photographs and videos). In addition, biological growths can threaten the health of those who frequent these places. The World Health Organization states that the most important effects are the increased prevalence of respiratory symptoms, allergies, asthma, and immune system disorders. This work aims to raise awareness of this fact and contribute to the identification of an automatic form of biological growth-type defects in images of buildings. To make this possible, we need a dataset of imaging building components with and without biological growths. Subsequently, deep learning methods are applied to allow the automatic identification of this type of defect in the images, and the results are analysed. A pre-trained VGG16 model was used. The dataset was annotated and divided into groups for training, validation, and testing. The model achieved an overall accuracy of 90%. This work demonstrates the potential of using Deep Learning (DL) in the maintenance and rehabilitation of urban infrastructures, highlighting the efficiency and sustainability of these processes and the importance of adjustments to ensure the stability of AI models. © 2025 The Author(s).

Abstract
Diabetic retinopathy is an eye disease that affects people with diabetes mellitus, causing lesions that affect the retina, leading to progressive vision loss. In Portugal, it is estimated that 1.5 million people between the ages of 20 and 79 have diabetes, a figure that is expected to rise in the coming years. This increase is also likely to raise the total number of people affected by diabetic retinopathy, who will need to be identified in the early stages of the disease to receive clinical treatment aimed at reducing the likelihood of visual impairment due to the disease. Detection and classification of the stage of severity is carried out by specialists using medical images of the retina and patients' clinical data. Fundus photographs are the standard for detecting and monitoring the progression of the disease, as they make it possible to see biomarkers that characterize the stages of the disease. The manual task of analyzing and tracing images is time-consuming and subjective, which can lead to interpretation errors. Artificial intelligence (AI) models, such as convolutional neural networks, have been proposed to aid specialists in medical image analysis tasks, some of which have already been approved for clinical use. To overcome the limitations of convolutional networks, new AI models have been proposed to develop computer vision applications, achieving promising results in image classification. The vision mamba model was recently introduced, which uses bidirectional state space to obtain an efficient visual representation. In this work, we evaluate the vision mamba model's ability to detect cases in the moderate and advanced stages of diabetic retinopathy in fundus photographs and compare its performance with models based on convolutional networks. As the best result, the model achieved a recall value of 0.95 in the APTOS dataset. © 2025 The Author(s).

Abstract
Digestive disorders can be signs of long-term conditions such as cancer, and as such, they should be treated seriously. Endoscopic exams of the gastrointestinal tract allow for the early detection of these conditions and facilitate effective treatment; these procedures have their effectiveness limited by variations in operator performance, due to human error. Support systems are desired to help specialists detect and diagnose pathologies in this type of exam. This work used a seldom utilized dataset, the ERS dataset, which contains 121,399 labeled images, to evaluate eight models from the EfficientNet family of architectures, as well as three models from the EfficientNetV2 iteration of this architecture, for the task of binary classification of endoscopic images. This work also compared their performance to four other widely used CNN architectures for the same task, along with the baseline results published by the authors of the dataset. Each model was evaluated in a 5-fold cross-validation procedure, following the same training protocol. The experiments have shown that the best-performing architecture was EfficientNetV2M, followed closely by EfficientNetB7, with the former achieving average accuracy and F1-Score values of, respectively, 82.24% and 88.15%. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2025.

Abstract
O image stitching é uma técnica que permite combinar múltiplas imagens, formando uma imagem única com um campo de visão amplo. No contexto da retinografia, essa técnica é crucial para capturar uma visão detalhada da retina, permitindo que exames mais amplos sejam realizados. Neste trabalho, é apresentado um método de costura de imagens de baixa resolução, utilizando pontos de bifurcação como características da imagem. O método proposto apresenta um aumento das correspondências obtidas em relação aos detectores da literatura, obtendo o resultado de RMSE com uma redução de aproximadamente 14% em comparação ao SIFT (24,29) e ORB (24,32). Além disso, o método proposto obteve um PSNR médio superior, atingindo 26,27 para imagens de glaucoma, 26,72 para imagens normais e 26,71 para imagens de pacientes suspeitos, enquanto os métodos baseados em SIFT e ORB apresentaram valores inferiores, confirmando a eficácia da abordagem proposta.

Abstract
Pneumonia is a respiratory condition caused by various microorganisms, including bacteria, viruses, fungi, and parasites. It manifests with symptoms such as coughing, chest pain, fever, breathing difficulties, and fatigue. Early and accurate detection is crucial for effective treatment, yet traditional diagnostic methods often fall short in reliability and speed. Chest X-rays have become widely used for detecting pneumonia; however, current approaches still struggle with achieving high accuracy and interpretability, leaving room for improvement. PneumoNet, an artificial intelligence assistant for X-ray pneumonia detection, is proposed in this work. The framework comprises (a) a new deep learning-based classification model for the detection of pneumonia, which expands on the AlexNet backbone for feature extraction in X-ray images and a new head in its final layers that is tailored for (X-ray) pneumonia classification. (b) GPT-Neo, a large language model, which is used to integrate the results and produce medical reports. The classification model is trained and evaluated on three publicly available datasets to ensure robustness and generalisability. Using multiple datasets mitigates biases from single-source data, addresses variations in patient demographics, and allows for meaningful performance comparisons with prior research. PneumoNet classifier achieves accuracy rates between 96.70% and 98.70% in those datasets.

Abstract
Inherited retinal diseases (IRDs) are rare and genetically diverse disorders that cause progressive vision loss and affect 1 in 3000 individuals worldwide. Their rarity and genetic variability pose a challenge for deep learning models due to the limited amount of data. Generative models offer a promising solution by creating synthetic data to improve training datasets. This study carried out a systematic literature review to investigate the use of generative models to augment data in IRDs and assess their impact on the performance of classifiers for these diseases. Following PRISMA 2020 guidelines, searches in four databases identified 32 relevant studies, 2 focused on IRD and the rest on other retinal diseases. The results indicate that generative models effectively augment small datasets. Among the techniques identified, Deep Convolutional Adversarial Generative Networks (DCGAN) and the Style-Based Generator Architecture of Generative Adversarial Networks 2 (StyleGAN2) were the most widely used. These architectures generated highly realistic and diverse synthetic data, often indistinguishable from real data, even for experts. The results highlight the need for more research into data generation in IRD to develop robust diagnostic tools and improve genetic studies by creating more comprehensive genetic repositories.