The advent of large–scale pre–training has produced powerful computer vision models, yet their effectiveness as feature extractors on highly specialized, out–of–distribution domains remains a critical issue. To fill this gap, we systematically compare the feature extraction capabilities of different, state–of–the–art architectures on the challenging task of dermoscopic skin–lesion classification—a domain substantially different from the natural images used during pre–training. We evaluate four distinct backbone architectures for feature extraction, encompassing both unimodal and multimodal models. Our selection includes EfficientNet B0 and Vision Transformer as representative unimodal models, along with Stable Diffusion Model and CLIP as state–of–the–art multimodal models. Using a subset of the ISIC archive, we assess the quality of extracted features via both linear probing and classification with a Multi Layer Perceptron. Our experiments suggest that all architectures produce highly effective features, achieving competitive performance. Notably, multimodal models, although pre–trained for tasks other than image classification, extract features that are remarkably competitive in this context. This work can serve as a comparative guide for the selection of feature extractors when tackling classification in specialized domains, such as medical imaging. The results highlight not only the generalizability of modern architectures, but also the surprising versatility of multimodal models as powerful feature extractors for interdisciplinary tasks.

This study investigates whether pre-trained Diffusion Models (DMs) inherently encode in a different way real and synthetic objects in an image within their latent representations. While DMs have achieved state-of-the-art performance in generative modeling and demonstrated rich semantic representations despite being trained solely with denoising objectives, their widespread accessibility raises critical concerns about detecting AI-generated content. Motivated by the intuition that a model must implicitly capture the distribution of real data to generate it, we hypothesized that real and synthetic objects are mapped to separable regions of the latent space. To test this hypothesis, we conduct a systematic analysis using a pre-trained DM to encode both real objects from the COCO and the Pascal VOC datasets and synthetic samples from DALL-E 3, Midjourney, and Stable Diffusion. We extract internal representations from different layers and evaluate their discriminative capacity through distance-based metrics and linear probing. Our findings reveal that decoder features, in particular, induce statistically meaningful representational differences between real and synthetic objects. The results show that these representations support high linear-probing accuracy under the evaluated experimental conditions, suggesting that pre-trained diffusion representations contain useful signals for distinguishing real and synthetic content.

Circle detection plays a pivotal role in computer vision, underpinning applications from industrial inspection and bioinformatics to autonomous driving. Traditional methods, however, often struggle with real-world complexities, as they demand extensive parameter tuning and adaptation across different domains. In this paper, we present the Synthetic Circle Dataset (SynCircle), a large synthetic image dataset designed to train a YOLO v10 network for circle detection. The YOLO v10 network, pre-trained solely on synthetic data, demonstrates remarkable off-the-shelf performance that surpasses conventional methods in various practical scenarios. Furthermore, we show that incorporating just a few labeled real images for fine-tuning can significantly boost performance, reducing the need for large annotated datasets. To promote reproducibility and streamline adoption, we publicly release both the trained YOLO v10 weights and the full SynCircle dataset.

Nystagmus is a condition characterized by involuntary, rapid and repetitive movement of the eyes, mainly caused by a dysfunction of the peripheral vestibular system located in the inner ear. It may also arise from the dysfunction of brain areas dedicated to the control of equilibrium and of the visual system. Precisely recognizing nystagmus may help in rapidly differentiating benign from potentially life-threatening diseases. In this paper, we propose a method, based on eye tracking, to easily process RGB videos captured with widely used devices (like smartphones and tablets) to recognize horizontal nystagmus and its direction. The proposed pipeline is designed to work in flexible conditions, without requiring specific recording configurations. We train and evaluate the system on real videos and report an AUROC of 88.7% for nystagmus detection and 76.5% for direction classification. These results support the feasibility of using consumer-grade video data for accurate and automatic screening of nystagmus.

Over the last decade, thanks to advances in sensor technology, imaging, and data analytics, the use of AIbased medical devices has seen an exponential increase, with massive adoption by hospitals and specialized laboratories. Among support devices for patients with limited or no residual mobility, and unable to interact to express their basic needs, artificial intelligence technologies offer promising solutions for augmentative and alternative communication (AAC), necessary for people with speech disorders, language impairment, and autism. This paper focuses on the development of a smart virtual keyboard for patients with reduced communication capabilities - specifically improving the AAC BrainControl Interface device, which uses an EEG helmet to detect users' brain activity - to help them interact with the outside world. By leveraging machine learning techniques for language generation, particularly using recurrent networks and large language models, it is possible to accurately predict user intent and improve their typing experience. Experimental results show the average number of interactions reduced by a factor of 2.66 compared to the original sequential key scanning method, which is extremely significant for locked-in patients.

The rapid progress of generative diffusion models has enabled the creation of synthetic images that are increasingly difficult to distinguish from real ones, raising concerns about authenticity, copyright, and misinformation. Existing supervised detectors often struggle to generalize across unseen generators, requiring extensive labeled data and frequent retraining. We introduce FRIDA (Fake-image Recognition and source Identification via Diffusion-features Analysis), a lightweight framework that leverages internal activations from a pre-trained diffusion model for deepfake detection and source generator attribution. A k-nearest-neighbor classifier applied to diffusion features achieves state-of-the-art cross-generator performance without fine-tuning, while a compact neural model enables accurate source attribution. These results show that diffusion representations inherently encode generator-specific patterns, providing a simple and interpretable foundation for synthetic image forensics.

Recently, generative models for images have garnered remarkable attention, due to their effective generalization ability and their capability to generate highly detailed and realistic content. Indeed, the success of generative networks ( e.g. , BigGAN, StyleGAN, Diffusion Models) has driven researchers to develop increasingly powerful models. Asa result, we have observed an unprecedented improvement in terms of both image resolution and realism, making generated images indistinguishable from real ones. In this work, we focus on a family of generative models known as Stable Diffusion Models (SDMs), which have recently emerged due to their ability to generate images in a multimodal setup ( i.e. , from a textual prompt) and have outperformed adversarial networks by learning to reverse a diffusion process. Given the complexity of these models that makes it hard to retrain them, researchers started to exploit pre-trained SDMs to perform downstream tasks ( e.g. , classification and segmentation), where semantics plays a fundamental role. In this context, understanding how well the model preserves semantic information maybe crucial to improve its performance. This paper presents an approach aimed at providing insights into the properties of a pre-trained SDM through the semantic lens. In particular, we analyze the features extracted by the U-Net within a SDM to explore whether and how the semantic information of an image is preserved in its internal representation. For this purpose, different distance measures are compared, and an ablation study is performed to select the layer (or combination of layers) of the U-Net that best preserves the semantic information. We also seek to understand whether semantics are preserved when the image undergoes simple transformations ( e.g. , rotation, flip, scale, padding, crop, and shift) and fora different number of diffusion denoising steps. To evaluate these properties, we consider popular benchmarks for semantic segmentation tasks ( e.g. , COCO, and PascalVOC). Our experiments suggest that the first encoder layer at 16x16 resolution effectively preserves semantic information. However, increasing inference steps (even fora minimal amount of noise) and applying various image transformations can affect the diffusion U-Net's internal feature representation. Additionally, we propose some examples taken from a video benchmark (DAVIS dataset), where we investigate if an object instance within a video preserves its internal representation even after several frames. Our findings suggest that the internal object representation remains consistent across multiple frames in a video, as long as the configuration changes are not excessive.

Retinal fundus imaging is crucial for diagnosing and monitoring eye diseases, which are often linked to systemic health conditions such as diabetes and hypertension. Current deep learning techniques often narrowly focus on segmenting retinal blood vessels, lacking a more comprehensive analysis and characterization of the retinal vascular system. This study fills this gap by proposing a novel, integrated approach that leverages multiple stages to accurately determine vessel paths and extract informative features from them. The segmentation of veins and arteries, achieved through a deep semantic segmentation network, is used by a newly designed algorithm to reconstruct individual vessel paths. The reconstruction process begins at the optic disc, identified by a localization network, and uses a recurrent neural network to predict the vessel paths at various junctions. The different stages of the proposed approach are validated both qualitatively and quantitatively, demonstrating robust performance. The proposed approach enables the extraction of critical features at the individual vessel level, such as vessel tortuosity and diameter. This work lays the foundation for a comprehensive retinal image evaluation, going beyond isolated tasks like vessel segmentation, with significant potential for clinical diagnosis.

The importance of kidney biopsy, a medical procedure in which a small tissue sample is extracted from the kidney for examination, is increasing due to the rising incidence of kidney disorders. This procedure helps diagnosing several kidney diseases which are cause of kidney function changes, as well as guiding treatment decisions, and evaluating the suitability of potential donor kidneys for transplantation. In this work, a deep learning system for the automatic segmentation of glomeruli in biopsy kidney images is presented. A novel cross-species transfer learning approach, in which a semantic segmentation network is trained on mouse kidney tissue images and then fine-tuned on human data, is proposed to boost the segmentation performance. The experiments conducted using two deep semantic segmentation networks, MobileNet and SegNeXt, demonstrated the effectiveness of the cross-species pre-training approach leading to an increased generalization ability of both models.

Online communication platforms have revolutionized interpersonal interactions by transcending geographical barriers. While facilitating connectivity, these platforms have introduced challenges such as overcoming linguistic differences and preventing spam and offensive content diffusion. This is particularly pertinent in the context of deliberative events, where online platforms could be used to extend the inclusion of citizens in democratic decision-making. In traditional deliberative events, human moderators and translators were used to facilitate conversation; however, the need for these figures imposed a limit on both the number of deliberative events that could be organized and the number of participants. In response, this paper proposes an automated moderator for deliberative events. The moderator is developed in Python for the online communication platform Discord and can be used, thanks to the integrated AI (Artificial Intelligence) tools, to automatically manage conversation agendas, prevent spam and inappropriate language, analyze the sentiment of the conversation, and translate messages into multiple languages. In particular, three classifiers, based on a pre-trained BERT (Bidirection Encoder Representations from Transformers), were fine-tuned for spam detection, toxic comments classification, and sentiment analysis. These allow the moderator to automatically detect and remove spam and offensive messages in different languages, send warnings to users, alert administrators, and, after repeated warnings, impose bans. Additionally, a built-in translator, based on Meta’s No Language Left Behind NLLB model, translates messages into five languages (Italian, English, French, German, and Polish). The developed bot was tested in a simulated deliberative event on a Discord server, demonstrating its ability to manage conversations and prevent linguistic abuse.

Whole genome sequencing of bacteria is important to enable strain classification. Using entire genomes as an input to machine learning (ML) models would allow rapid classification of strains while using information from multiple genetic elements. We developed a “bag-of-words” approach to encode, using SentencePiece or k-mer tokenization, entire bacterial genomes and analyze these with ML. Initial model selection identified SentencePiece with 8,000 and 32,000 words as the best approach for genome tokenization. We then classified in Neisseria meningitidis genomes the capsule B group genotype with 99.6% accuracy and the multifactor invasive phenotype with 90.2% accuracy, in an independent test set. Subsequently, in silico knockouts of 2,808 genes confirmed that the ML model predictions aligned with our current understanding of the underlying biology. To our knowledge, this is the first ML method using entire bacterial genomes to classify strains and identify genes considered relevant by the classifier.

The ambition to create increasingly realistic images has driven researchers to develop increasingly powerful models, capable of generalizing and generating high-resolution images, even in a multimodal setup (e.g., from textual input). Among the most recent generative networks, Stable Diffusion Models (SDMs) have achieved state-of-the-art showing great generative capabilities but also a high degree of complexity, both in terms of training and interpretability. Indeed, the impressive generalization capability of pre-trained SDMs has pushed researchers to exploit their internal representation to perform downstream tasks (e.g., classification and segmentation). Understanding how well the model preserves semantic information is fundamental to improve its performance. Our approach, namely Diff-Props, analyses the features extracted from the U-Net within Stable Diffusion Model to unveil how Stable Diffusion retains semantic information of an image in a pre-trained setup. Exploiting a set of different distance metrics, Diff-Props aims to analyse how features at different depths contribute to preserving the meaning of the objects in the image.

In an era characterized by rapid technological advancement and rising customer expectations, accurate ticket classification in banking customer service emerges as a critical necessity. In this context, we designed a comprehensive ticket classification pipeline, leveraging a real-world dataset comprising 4,243 chat-based user requests, classified into ten distinct classes, provided by MPS Bank. Our approach proposes a complete data processing pipeline with an exploration of two text classification methodologies: BERT (Bidirectional Encoder Representations from Transformers) and TF-IDF (Term Frequency - Inverse Document Frequency) with SVM (Support Vector Machine). The experiments highlight that both models have considerable potential, promising substantial improvements in the operational efficiency of customer support, ultimately increasing the overall quality of service.

Fluorescence imaging plays a crucial role in studying biological processes and materials across various industrial applications. However, the manual analysis of fluorescence images is time-consuming and prone to errors. To address these challenges, we propose novel machine learning-based approaches to enhance the VIDAS® device - an automated immunoassay system employed for medical condition detection. The current setup utilizes a photodiode for luminescence capture, which exhibits limitations when applied to spatially distributed signals. To overcome this limitation, we explore the use of a CMOS sensor to capture two-dimensional images of cuvettes, enabling a more comprehensive analysis of the system. Our proposed solution involves generating reconstructed images that rectify potential defects, leading to improved and unbiased fluorescence estimation. Through extensive experimentation, we demonstrate that employing the reconstructed images enables more accurate measurements, particularly in the presence of defects. Our methodology encompasses deep learning and semantic segmen-tation techniques, allowing robust fluorescence image analysis.

Recognizing skin cancer in time could greatly increase patients' chances of recovery. For this reason, in recent years, numerous decision support systems have been proposed to help dermatologists in this diagnosis. These systems are generally based on Convolutional Neural Networks and are used for both segmentation and classification of lesions. Although their main goal is to correctly recognize the lesions' type, the preliminary segmentation step has been shown to increase the performance of the classifier. In fact, this is not surprising because physicians also use information on the shape of the lesion to make a diagnosis. Thanks to the ISIC archive, a huge number of skin lesion images, along with the corresponding metadata (type, position, dimension, etc.), are publicly available to train a deep neural network, but, unfortunately, only a small fraction of them are labeled for segmentation. To overcome this limitation, in this paper, a weak supervised approach is proposed to extract the segmentation label maps from the entire ISIC archive. Moreover, to demonstrate the quality of the proposed approach, the generated supervisions were first compared with those available in ISIC and, then, used to train a segmentation network, whose performance was evaluated against that obtained using only the small set of ISIC label maps. To foster reproducibility and to promote future research in lesion segmentation and classification, the generated ISIC Weak Segmentation Map (ISIC_WSM) dataset has been released. As far as we know, this is the first dataset that contains segmentation supervisions for clinical images of skin lesions.

Skin cancer is a serious public health problem with a sharply increasing incidence in recent years, which has a major impact on quality of life and can be disfiguring or even fatal. Deep learning techniques can be used to analyze dermoscopic images, resulting in automated systems that can improve the clinical confidence of the diagnosis – also avoiding unnecessary surgery – help clinicians objectively communicate its outcome, reduce errors related to human fatigue, and cut costs affecting the health system. In this chapter, we present an entire pipeline to analyze skin lesion images in order to distinguish nevi from melanomas, also integrating patient clinical data to reach a diagnosis. Furthermore, to make our artificial intelligence tool explainable for both clinicians and patients, dermoscopic images are further processed to obtain their segmented counterparts, where the lesion contour is easily observable, and saliency maps, highlighting the areas of the lesion that prompted the classifier to make its decision. Experimental results are promising and have been positively evaluated by human experts.

In this paper, we use Generative Adversarial Networks (GANs) to synthesize high-quality retinal images along with the corresponding semantic label-maps, instead of real images during training of a segmentation network. Different from other previous proposals, we employ a two-step approach: first, a progressively growing GAN is trained to generate the semantic label-maps, which describes the blood vessel structure (i.e., the vasculature); second, an image-to-image translation approach is used to obtain realistic retinal images from the generated vasculature. The adoption of a two-stage process simplifies the generation task, so that the network training requires fewer images with consequent lower memory usage. Moreover, learning is effective, and with only a handful of training samples, our approach generates realistic high-resolution images, which can be successfully used to enlarge small available datasets. Comparable results were obtained by employing only synthetic images in place of real data during training. The practical viability of the proposed approach was demonstrated on two well-established benchmark sets for retinal vessel segmentation-both containing a very small number of training samples-obtaining better performance with respect to state-of-the-art techniques.