Accurate segmentation of Multiple Sclerosis (MS) lesions remains a critical challenge in medical image analysis due to their small size, irregular shape, and sparse distribution. Despite recent progress in vision foundation models — such as SAM and its medical variant MedSAM — these models have not yet been explored in the context of MS lesion segmentation. Moreover, their reliance on manually crafted prompts and high inference-time computational cost limits their applicability in clinical workflows, especially in resource-constrained environments. In this work, we introduce a novel training-time framework for effective and efficient MS lesion segmentation. Our method leverages SAM solely during training to guide a prompt learner that automatically discovers task-specific embeddings. At inference, SAM is replaced by a lightweight convolutional aggregator that maps the learned embeddings directly into segmentation masks—enabling fully automated, low-cost deployment. We show that our approach significantly outperforms existing specialized methods on the public MSLesSeg dataset, establishing new performance benchmarks in a domain where foundation models had not previously been applied. To assess generalizability, we also evaluate our method on pancreas and prostate segmentation tasks, where it achieves competitive accuracy while requiring an order of magnitude fewer parameters and computational resources compared to SAM-based pipelines. By eliminating the need for foundation models at inference time, our framework enables efficient segmentation without sacrificing accuracy. This design bridges the gap between large-scale pretraining and real-world clinical deployment, offering a scalable and practical solution for MS lesion segmentation and beyond. Code is available at https://github.com/perceivelab/MS-SAM-LESS .

CLIP has revolutionized zero-shot learning by enabling task generalization without fine-tuning. While prompting techniques like CoOp and CoCoOp enhance CLIP’s adaptability, their effectiveness in Federated Learning (FL) remains an open challenge. Existing federated prompt learning approaches, such as FedCoOp and FedTPG, improve performance but face generalization issues, high communication costs, and reliance on a central server, limiting scalability and privacy.We propose Zero-shot Decentralized Federated Learning (ZeroDFL), a fully decentralized framework that enables zero-shot adaptation across distributed clients without a central coordinator. ZeroDFL employs an iterative prompt-sharing mechanism, allowing clients to optimize and exchange textual prompts to enhance generalization while drastically reducing communication overhead.We validate ZeroDFL on nine diverse image classification datasets, demonstrating that it consistently outperforms—or remains on par with—state-of-the-art federated prompt learning methods. More importantly, ZeroDFL achieves this performance in a fully decentralized setting while reducing communication overhead by 118× compared to FedTPG. These results highlight that our approach not only enhances generalization in federated zero-shot learning but also improves scalability, efficiency, and privacy preservation—paving the way for decentralized adaptation of large vision-language models in real-world applications. Code is available at: https://github.com/perceivelab/ZeroDFL

Foundation models have transformed multimedia analysis by enabling robust and transferable representations across diverse modalities and tasks. However, their static deployment conflicts with growing societal and regulatory demands-particularly the need to unlearn specific data upon request, as mandated by privacy frameworks such as the GDPR. Traditional unlearning approaches, including retraining, activation editing, or distillation, are often computationally expensive, fragile, and ill-suited for real-time or continuously evolving systems. In this paper, we propose a paradigm shift: rethinking unlearning not as a retroactive intervention but as a built-in capability. We introduce a prompt-based learning framework that unifies knowledge acquisition and removal within a single training phase. Rather than encoding information in model weights, our approach binds class-level semantics to dedicated prompt tokens. This design enables instant unlearning simply by removing the corresponding prompt-without retraining, model modification, or access to original data. Experiments demonstrate that our framework preserves predictive performance on retained classes while effectively erasing forgotten ones. Beyond utility, our method exhibits strong privacy and security guarantees: it is resistant to membership inference attacks, and prompt removal prevents any residual knowledge extraction, even under adversarial conditions. This ensures compliance with data protection principles and safeguards against unauthorized access to forgotten information, making the framework suitable for deployment in sensitive and regulated environments. Overall, by embedding removability into the architecture itself, this work establishes a new foundation for designing modular, scalable and ethically responsive AI models.

Understanding and explaining the behavior of machine learning models is essential for building transparent and trustworthy AI systems. We introduce DEXTER, a data-free framework that employs diffusion models and large language models to generate global, textual explanations of visual classifiers. DEXTER operates by optimizing text prompts to synthesize class-conditional images that strongly activate a target classifier. These synthetic samples are then used to elicit detailed natural language reports that describe class-specific decision patterns and biases. Unlike prior work, DEXTER enables natural language explanation about a classifier's decision process without access to training data or ground-truth labels. We demonstrate DEXTER's flexibility across three tasks—activation maximization, slice discovery and debiasing, and bias explanation—each illustrating its ability to uncover the internal mechanisms of visual classifiers. Quantitative and qualitative evaluations, including a user study, show that DEXTER produces accurate, interpretable outputs. Experiments on ImageNet, Waterbirds, CelebA, and FairFaces confirm that DEXTER outperforms existing approaches in global model explanation and class-level bias reporting. Code is available at https://github.com/perceivelab/dexter.

We present DiffExplainer , a novel framework that, leveraging language-vision models, enables multimodal global explainability. DiffExplainer employs diffusion models conditioned on optimized text prompts, synthesizing images that maximize class outputs and hidden features of a classifier, thus providing a visual tool for explaining decisions. Moreover, the analysis of generated visual descriptions allows for automatic identification of biases and spurious features, as opposed to traditional methods that often rely on manual intervention. The cross-modal transferability of language-vision models also enables the possibility to describe decisions in a more human-interpretable way, i.e., through text. We conduct comprehensive experiments demonstrating the effectiveness of DiffExplainer on (1) the generation of high-quality images explaining model decisions, surpassing existing activation maximization methods, and (2) the automated identification of biases and spurious features. • Leverages diffusion models to generate images explaining classifier decisions. • Enables bias and spurious feature detection without manual intervention. • Outperforms activation maximization methods in image quality and feature analysis. • Enables specific model analysis by the use of fixed prompts.

We propose wake-sleep consolidated learning (WSCL), a learning strategy leveraging complementary learning system (CLS) theory and the wake-sleep phases of the human brain to improve the performance of deep neural networks (DNNs) for visual classification tasks in continual learning (CL) settings. Our method learns continually via the synchronization between distinct wake and sleep phases. During the wake phase, the model is exposed to sensory input and adapts its representations, ensuring stability through a dynamic parameter freezing mechanism and storing episodic memories in a short-term temporary memory (similar to what happens in the hippocampus). During the sleep phase, the training process is split into nonrapid eye movement (NREM) and rapid eye movement (REM) stages. In the NREM stage, the model's synaptic weights are consolidated using replayed samples from the short-term and long-term memory and the synaptic plasticity mechanism is activated, strengthening important connections and weakening unimportant ones. In the REM stage, the model is exposed to previously-unseen realistic visual sensory experience, and the dreaming process is activated, which enables the model to explore the potential feature space, thus preparing synapses for future knowledge. We evaluate the effectiveness of our approach on four benchmark datasets: CIFAR-10, CIFAR-100, Tiny-ImageNet, and FG-ImageNet. In all cases, our method outperforms the baselines and prior work, yielding a significant performance gain on continual visual classification tasks. Furthermore, we demonstrate the usefulness of all processing stages and the importance of dreaming to enable positive forward transfer (FWT). The code is available at: https://github.com/perceivelab/wscl.

In this paper, we propose ViTO, a novel knowledge distillation strategy that aims to convert a CNN model into a transformer-based counterpart that incorporates the advantages of transformers while retaining or improving its inductive bias. Our approach is based on a two-level transformer architecture that includes an inner model for learning visual representations and an outer model that aims to match the teacher’s predictions through autoregression. Specifically, given an image in a batch, the outer model classifies the image by using, in addition to the image’s visual properties, also the predictions it has made on images previously seen within the same batch. The effect of this strategy is to allow the transformer to estimate self- and cross-attention across all input batch images to learn autoregressively intra-class and inter-class correlations.We experimentally validate ViTO on several standard benchmarks obtaining better performance than existing knowledge distillation strategies on transformers. Furthermore, our distilled transformer-based model shows better robustness properties than standard vision transformers, demonstrating the effectiveness of our proposed distillation strategy.

We present Saliency-driven Experience Replay - SER - a biologically-plausible approach based on replicating human visual saliency to enhance classification models in continual learning settings. Inspired by neurophysiological evidence that the primary visual cortex does not contribute to object manifold untangling for categorization and that primordial saliency biases are still embedded in the modern brain, we propose to employ auxiliary saliency prediction features as a modulation signal to drive and stabilize the learning of a sequence of non-i.i.d. classification tasks. Experimental results confirm that SER effectively enhances the performance (in some cases up to about twenty percent points) of state-of-the-art continual learning methods, both in class-incremental and task-incremental settings. Moreover, we show that saliency-based modulation successfully encourages the learning of features that are more robust to the presence of spurious features and to adversarial attacks than baseline methods. Code is available at: https://github.com/perceivelab/SER

In the medical field, multi-center collaborations are often sought to yield more generalizable findings by leveraging the heterogeneity of patient and clinical data. However, recent privacy regulations hinder the possibility to share data, and consequently, to come up with machine learning-based solutions that support diagnosis and prognosis. Federated learning (FL) aims at sidestepping this limitation by bringing AI-based solutions to data owners and only sharing local AI models, or parts thereof, that need then to be aggregated. However, most of the existing federated learning solutions are still at their infancy and show several shortcomings, from the lack of a reliable and effective aggregation scheme able to retain the knowledge learned locally to weak privacy preservation as real data may be reconstructed from model updates. Furthermore, the majority of these approaches, especially those dealing with medical data, relies on a centralized distributed learning strategy that poses robustness, scalability and trust issues. In this paper we present a federated learning strategy, FedER, that, exploiting experience replay and generative adversarial concepts, effectively integrates features from local nodes, providing models able to generalize across multiple datasets while maintaining privacy. FedER is tested on two tasks — tuberculosis and melanoma classification — using multiple datasets in order to simulate realistic non-i.i.d. medical data scenarios. Results show that our approach achieves performance comparable to standard (non-federated) learning and significantly outperforms state-of-the-art federated methods. Remarkably, we also observe that FedER enables any node model to be used as a global federation model. Indeed, the experience replay strategy with privacy-preserving synthetic data allows all node models to converge to reach the same optimum without the need of a single shared model. Code is available at https://github.com/perceivelab/FedER.

The quantification of stenosis severity from X-ray catheter angiography is a challenging task. Indeed, this requires to fully understand the lesion's geometry by analyzing dynamics of the contrast material, only relying on visual observation by clinicians. To support decision making for cardiac intervention, we propose a hybrid CNN-Transformer model for the assessment of angiography-based non-invasive fractional flow-reserve (FFR) and instantaneous wave-free ratio (iFR) of intermediate coronary stenosis. Our approach predicts whether a coronary artery stenosis is hemodynamically significant and provides direct FFR and iFR estimates. This is achieved through a combination of regression and classification branches that forces the model to focus on the cut-off region of FFR (around 0.8 FFR value), which is highly critical for decision-making. We also propose a spatio-temporal factorization mechanisms that redesigns the transformer's self-attention mechanism to capture both local spatial and temporal interactions between vessel geometry, blood flow dynamics, and lesion morphology. The proposed method achieves state-of-the-art performance on a dataset of 778 exams from 389 patients. Unlike existing methods, our approach employs a single angiography view and does not require knowledge of the key frame; supervision at training time is provided by a classification loss (based on a threshold of the FFR/iFR values) and a regression loss for direct estimation. Finally, the analysis of model interpretability and calibration shows that, in spite of the complexity of angiographic imaging data, our method can robustly identify the location of the stenosis and correlate prediction uncertainty to the provided output scores.

In this paper, we propose a deep model based on graph convolutional networks for emotion recognition using EEG data. The model encodes spatial and temporal features of EEG channels and learns relationships between nodes through a self-attention mechanism, capturing spatio-temporal synchrony in brain regions. Experimental results show that our model outperforms existing approaches, with the attention mechanism contributing significantly to classification accuracy. In particular, the attention scores provide insights into how EEG channels influence each other at different times, revealing spatio-temporal patterns of brain connectivity related to emotions.

Video saliency prediction has recently attracted attention of the research community, as it is an upstream task for several practical applications. However, current solutions are particurly computationally demanding, especially due to the wide usage of spatio-temporal 3D convolutions. We observe that, while different model architectures achieve similar performance on benchmarks, visual variations between predicted saliency maps are still significant. Inspired by this intuition, we propose a lightweight model that employs multiple simple heterogeneous decoders and adopts several practical approaches to improve accuracy while keeping computational costs low, such as hierarchical multi-map knowledge distillation, multi-output saliency prediction, unlabeled auxiliary datasets and channel reduction with teacher assistant supervision. Our approach achieves saliency prediction accuracy on par or better than state-of-the-art methods on DFH1K, UCF-Sports and Hollywood2 benchmarks, while enhancing significantly the efficiency of the model.

In Continual Learning (CL), a neural network is trained on a stream of data whose distribution changes over time. In this context, the main problem is how to learn new information without forgetting old knowledge (i.e., Catastrophic Forgetting). Most existing CL approaches focus on finding solutions to preserve acquired knowledge, so working on the "past" of the model. However, we argue that as the model has to continually learn new tasks, it is also important to put focus on the "present" knowledge that could improve following tasks learning. In this paper we propose a new, simple, CL algorithm that focuses on solving the current task in a way that might facilitate the learning of the next ones. More specifically, our approach combines the main data stream with a secondary, diverse and uncorrelated stream, from which the network can draw auxiliary knowledge. This helps the model from different perspectives, since auxiliary data may contain useful features for the current and the next tasks and incoming task classes can be mapped onto auxiliary classes. Furthermore, the addition of data to the current task is implicitly making the classifier more robust as we are forcing the extraction of more discriminative features. Our method can outperform existing state-of-the-art models on the most common CL Image Classification benchmarks.

Abstract In this work, we propose a 3D fully convolutional architecture for video saliency prediction that employs hierarchical supervision on intermediate maps (referred to as conspicuity maps) generated using features extracted at different abstraction levels. We provide the base hierarchical learning mechanism with two techniques for domain adaptation and domain-specific learning. For the former, we encourage the model to unsupervisedly learn hierarchical general features using gradient reversal at multiple scales, to enhance generalization capabilities on datasets for which no annotations are provided during training. As for domain specialization, we employ domain-specific operations (namely, priors, smoothing and batch normalization) by specializing the learned features on individual datasets in order to maximize performance. The results of our experiments show that the proposed model yields state-of-the-art accuracy on supervised saliency prediction. When the base hierarchical model is empowered with domain-specific modules, performance improves, outperforming state-of-the-art models on three out of five metrics on the DHF1K benchmark and reaching the second-best results on the other two. When, instead, we test it in an unsupervised domain adaptation setting, by enabling hierarchical gradient reversal layers, we obtain performance comparable to supervised state-of-the-art. Source code, trained models and example outputs are publicly available at https://github.com/perceivelab/hd2s .

In this paper we define a deep multi-stage architecture for automated landmarking of craniomaxillofacial (CMF) CT images. Our model is composed of three subnetworks that first localize, on reduced-resolution images, areas where landmarks may be found and then refine the search, at full-resolution scale, through a hierarchical structure aiming at increasing the granularity of the investigated region. The multi-stage pipeline is designed to deal with full resolution data and does not require any additional pre-processing step to reduce search space, as opposed to existing methods that can be only adopted for searching landmarks located in well-defined anatomical structures (e.g., mandibles). The automated landmarking system is tested on identifying landmarks located in several CMF regions, achieving an average error of 0.8 mm, significantly lower than expert readings. The proposed model also outperforms baselines and is on par with existing models that employ additional upstream segmentation, on state-of-the-art benchmarks.