Publications

Per stare al passo con la natura in continua evoluzione dei dati, i moderni sistemi di intelligenza artificiale richiedono frequenti e costosi riaddestramenti su tutti gli esempi già visti, per evitare il fenomeno del catastrophic forgetting. Questa esigenza ha stimolato un crescente interesse verso il Continual Learning (CL), in cui i modelli apprendono in modo incrementale dai flussi di dati conservando al contempo le conoscenze acquisite. Tuttavia, nonostante i notevoli progressi, permane una fonte di incertezza spesso trascurata: la presenza di etichette rumorose o non affidabili. Le reti neurali profonde devono gran parte del loro successo a grandi dataset puliti, ma in contesti reali e dinamici tali condizioni ideali sono rare. Questo pone una domanda cruciale: come può un sistema continuare ad apprendere efficacemente quando la supervisione stessa è imperfetta? Per affrontare il problema dell’apprendimento con etichette rumorose in scenari incrementali, proponiamo Alternate Experience Replay (AER), una strategia che alterna fasi di apprendimento e “dimenticanza” del buffer, favorendo la separazione tra campioni puliti e rumorosi. In questo modo, il modello affina progressivamente le proprie rappresentazioni interne limitando la propagazione del rumore. A complemento, introduciamo Asymmetric Balanced Sampling (ABS), un meccanismo di campionamento che bilancia dinamicamente la conservazione di esempi puliti e complessi durante l’aggiornamento del buffer. La combinazione di questi approcci migliora la robustezza e la stabilità del modello, dimostrando come anche semplici meccanismi di replay possano ridurre l’impatto della supervisione imperfetta. La robustezza da sola, tuttavia, non basta. Una seconda sfida riguarda la capacità del modello di comporre e riutilizzare le conoscenze tra compiti diversi, proprietà nota come composizionalità. In questa tesi, mostriamo come rappresentazioni modulari e prospettive di ottimizzazione di secondo ordine possano favorire questa capacità. Introduciamo due paradigmi complementari: Incremental Task Arithmetic (ITA), che ottimizza ciascun modello addestrato su un singolo task individualmente, e Incremental Ensemble Learning (IEL), che ottimizza direttamente la loro composizione. Insieme, questi approcci permettono di combinare componenti apprese progressivamente in sistemi che non solo resistono alla dimenticanza, ma generalizzano per costruzione, adattandosi a nuove e inattese combinazioni di task di apprendimento. Queste due prospettive, robustezza alla supervisione rumorosa e adattamento composizionale, delineano un modello di Continual Learning resiliente e strutturato. Attraverso analisi empiriche approfondite, mostriamo come i meccanismi intrinseci delle reti neurali possano essere sfruttati per sviluppare modelli incrementali più affidabili e robusti. La tesi contribuisce alla ricerca nel campo del Continual Learning, migliorando la robustezza sia in presenza di supervisione rumorosa sia attraverso la composizione modulare della conoscenza. Forniamo una panoramica dello stato dell’arte, approfondimenti metodologici e studi sperimentali accurati su task incrementali complessi, con l’obiettivo di favorire lo sviluppo futuro di sistemi di apprendimento più adattivi e affidabili.

Recent advances in Virtual Try-On (VTON) and Virtual Try-Off (VTOFF) have greatly improved photo-realistic fashion synthesis and garment reconstruction. However, existing datasets remain static, lacking instruction-driven editing for controllable and interactive fashion generation. In this work, we introduce the Dress Editing Dataset (Dress-ED), the first large-scale benchmark that unifies VTON, VTOFF, and text-guided garment editing within a single framework. Each sample in Dress-ED includes an in-shop garment image, the corresponding person image wearing the garment, their edited counterparts, and a natural-language instruction of the desired modification. Built through a fully automated multimodal pipeline that integrates MLLM-based garment understanding, diffusion-based editing, and LLM-guided verification, Dress-ED comprises over 146k verified quadruplets spanning three garment categories and seven edit types, including both appearance (e.g., color, pattern, material) and structural (e.g., sleeve length, neckline) modifications. Based on this benchmark, we further propose a unified multimodal diffusion framework that jointly reasons over linguistic instructions and visual garment cues, serving as a strong baseline for instruction-driven VTON and VTOFF. Dataset and code available at this link: https://github.com/aimagelab/Dress-ED.

We present a predictive maintenance approach for industrial production lines based on multivariate segment-wise time-series analysis. To address the high cost of collecting anomalous samples, we propose a novelty detection framework in which a transformer autoencoder is trained in a semi-supervised fashion exclusively on nominal sequences, and anomaly scores are derived from reconstruction error at test time. We introduce a set of learnable “compression tokens” into the transformer encoder; these tokens serve as the bottleneck from which the decoder reconstructs the input. We compare this model against an MLP-based autoencoder baseline; the results show that the novelty-detection model remains strong, with near-perfect performance under time-aware and device-aware validation, which are the conditions that most faithfully simulate deployment.

Intraoral photographs (IOPs) provide a low-cost record of tooth appearance, alignment, soft tissue, and occlusal relationships. We present IOP-Compass, a dataset, benchmark, and annotation resource built on 1,000 patients from the Bite2Text collection. It comprises 5,000 standardized clinical photographs, five views per patient, with view labels and expert-verified tooth-instance masks carrying FDI numbers, together with frozen patient-disjoint splits. The dataset was produced through a browser-based human-in-the-loop annotation platform that we also released. Using IOP-Compass, we benchmark a pipeline representative of the current literature for view classification, region-of-interest extraction, and FDI-aware tooth instance segmentation, providing reference results and ablations across alternative pipeline components. View classification is near-saturated on the benchmark, while FDI-aware instance segmentation remains the main challenge. Both the dataset and code are publicly released.

Reproducibility policies promise "checkable" medical-imaging science, yet many submissions still ship unverifiable artifacts. Our analysis of 3722 MICCAI papers shows code-linking rising from 51.8% (2021) to 72.5% (2025), but ~13% of linked repositories are inaccessible or empty. We present paper-snitch, a reviewer-facing decision-support tool that turns these signals into an evidence-grounded report. Paper-snitch parses PDFs, resolves and sanity-checks repositories, and applies policy-aware checklists aligned with MICCAI expectations, producing a review-time verifiability score decomposed into interpretable sub-scores plus criterion-linked excerpts and artifacts reviewers can inspect. It never executes untrusted code or attempts GPU-heavy reproduction, focusing instead on bounded, verifiable checks. We compare paper-snitch on 100 randomly sampled MICCAI 2025 papers with human annotators using shared evaluation criteria, indicating that automated, bounded checks can scale reproducibility screening while keeping final decisions with reviewers.

Down syndrome (DS) represents one of the most common genetic disorders attributable to a partial or complete trisomy of chromosome 21 that affects about 1 in 700 individuals at birth. The diagnosis of Alzheimer's Disease (AD)-correlated cognitive decline in this population requires new approaches and new biomarkers that comprehensively assess health status and early cognitive decline. In this observational study, we explored for the first time the relation of IL-18, a cytokine member of IL-1 family involved in both innate and acquired immune responses, with DS associated cognitive decline. We observed that plasma total IL-18, in subjects with DS over 35 with and without AD-related cognitive decline, and plasma concentrations of its binding protein in subjects with DS (19-35 years) were correlated with lower plasma concentrations of Transforming Growth Factor (TGF-beta 1), which are linked to an increased rate of cognitive decline in adults with DS. In addition, we found a significant association between low baseline concentrations of Free IL-18, the active form of the cytokine, and an increased rate of cognitive decline at 12 months, calculated as delta of the Test for Severe Impairment (dTSI), in individuals with DS (19-35 years). Finally, we demonstrated a reduction of Free IL-18/TNF-alpha ratio, considered as a new possible double biomarker, in both young and older adult DS subjects without AD-related cognitive decline (area under the receiver operating curve (AUC) was 0.82 and 0.71, respectively), suggesting the advantage of the composite biomarkers in the discrimination of patients from healthy people over single biomarkers.

Diffusion Transformers (DiTs) and related flow-based architectures are now among the strongest text-to-image generators, yet the internal mechanisms through which prompts shape image semantics remain poorly understood. In this work, we study massive activations: a small subset of hidden-state channels whose responses are consistently much larger than the rest. We show that, despite their sparsity, these few channels effectively draw the whole picture, in three complementary senses. First, they are functionally critical: a controlled disruption probe that zeroes the massive channels causes a sharp collapse in generation quality, while disrupting an equally-sized set of low-statistic channels has marginal effect. Second, they are spatially organized: restricting image-stream tokens to massive channels and clustering them yields coherent partitions that closely align with the main subject and salient regions, exposing a structured spatial code hidden inside an apparently outlier-like subspace. Third, they are transferable: transporting massive activations from one prompt-conditioned trajectory into another, shifts the final image toward the source prompt while preserving substantial content from the target, producing localized semantic interpolation rather than unstructured pixel blending. We exploit this property in two use cases: text-conditioned and image-conditioned semantic transport, where massive activations transport enables prompt interpolation and subject-driven generation without any additional training. Together, these results recast massive activations not as activation anomalies, but as a sparse prompt-conditioned carrier subspace that organizes and controls semantic information in modern DiT models.

Male infertility is a significant yet under-addressed global health condition, and testicular ultrasound (US) plays a central role in its diagnostic evaluation. We introduce and publicly release TesticulUS-Real, the first multicenter testicular US segmentation dataset, addressing the absence of annotated public benchmarks for this anatomy. The dataset comprises 1,053 real ultrasound images acquired from two independent clinical institutions, with expert segmentation masks obtained through a standardized annotation and consensus review protocol. Leveraging this resource, we define an open-organ segmentation protocol to evaluate how models trained on existing multi-organ US data transfer to a previously unseen anatomical target. Beyond the dataset release, we conduct a broad cross-organ segmentation generalization study on ultrasound data. Using the UUSIC benchmark, we evaluate generalization across five anatomical regions and independent acquisition centers, comparing task-specific segmentation models, generalization-oriented ultrasound methods, and SAM-based foundation models under fully automatic inference. Alongside this benchmark, we introduce Cond-UNet, a lightweight conditional U-Net that combines Feature-wise Linear Modulation (FiLM) with our newly proposed shared attention conditioning (SAC) to obtain adaptive organ-aware representations. Experiments show that Cond-UNet achieves the best average cross-organ generalization performance across the UUSIC organs while using fewer parameters and lower computational cost than foundation-model alternatives. In the open-organ setting, the proposed testicular dataset enables a direct analysis of how different model families behave when facing an unseen ultrasound anatomy, highlighting the role of large-scale pretraining for foundation models and the robustness of organ-aware conditioning in lightweight architectures. The code is publicly released at https://github.com/AImageLab-zip/US_Cond-UNet, and the dataset at https://ditto.ing.unimore.it/testiculus/.

Object-centric models inspired by DETR have become the dominant paradigm for open-vocabulary video instance segmentation (OV-VIS). While recent efforts have reduced the computational cost of pixel decoding, textual modality fusion, and object decoding to make these architectures more suitable for mobile devices, real-time on-device inference at high frame rates remains an open challenge. In this paper, we introduce SegFS, a dual-stream fast-slow framework that significantly improves efficiency without sacrificing accuracy. On sparse keyframes, an open-vocabulary object-based model predicts instance-level representations. These representations are then projected back into the backbone feature space to condition a lightweight fast network, which efficiently relocalizes and segments the instances in subsequent frames. By shifting instance propagation from object decoding to feature-space conditioning, our approach decouples multimodal semantic understanding from dense mask prediction and enables efficient temporal propagation. The proposed fast branch achieves up to 14x lower latency than the mobile-oriented MOBIUS model, while maintaining competitive segmentation performance on standard OV-VIS benchmarks.