Publications by Vittorio Pipoli

Multimodal Large Language Models (MLLMs) have achieved impressive performance across a variety of vision–language tasks. However, their internal working mechanisms remain largely underexplored. In his work, we introduce FG-TRACER, a framework designed to analyze the information flow between visual and textual modalities in MLLMs in free-form generation. Notably, our numerically stabilized computational method enables the first systematic analysis of multimodal information flow in underexplored domains such as image captioning and chain-of-thought (CoT) reasoning. We apply FG-TRACER to two state-of-the-art MLLMs—LLaMA 3.2-Vision and LLaVA 1.5—across three vision–language benchmarks—TextVQA, COCO 2014, and ChartQA—and we conduct a word-level analysis of multimodal integration. Our findings uncover distinct patterns of multimodal fusion across models and tasks, demonstrating that fusion dynamics are both model- and task-dependent. Overall, FG-TRACER offers a robust methodology for probing the internal mechanisms of MLLMs in free-form settings, providing new insights into their multimodal reasoning strategies. Our source code is publicly available at https://anonymous.4open.science/r/FG-TRACER-CB5A/.

CNNs and Transformer-based architectures are recently dominating the field of 3D medical segmentation. While CNNs face limitations in the local receptive field, Transformers require significant memory and data, making them less suitable for analyzing large 3D medical volumes. Consequently, fully convolutional network models like U-Net are still leading the 3D segmentation scenario. Although efforts have been made to reduce the Transformers computational complexity, such optimized models still struggle with content-based reasoning. This paper examines Mamba, a Recurrent Neural Network (RNN) based on State Space Models (SSMs), which achieves linear complexity and has outperformed Transformers in long-sequence tasks. Specifically, we assess Mamba’s performance in 3D medical segmentation using three widely recognized and commonly employed datasets and propose architectural enhancements to improve its segmentation effectiveness by mitigating the primary shortcomings of existing Mamba-based solutions.

Brain tumor segmentation is a crucial task in medical imaging that involves the integrated modeling of four distinct imaging modalities to identify tumor regions accurately. Unfortunately, in real-life scenarios, the full availability of such four modalities is often violated due to scanning cost, time, and patient condition. Consequently, several deep learning models have been developed to address the challenge of brain tumor segmentation under conditions of missing imaging modalities. However, the majority of these models have been evaluated using the 2018 version of the BraTS dataset, which comprises only $285$ volumes. In this study, we reproduce and extensively analyze the most relevant models using BraTS2023, which includes 1,250 volumes, thereby providing a more comprehensive and reliable comparison of their performance. Furthermore, we propose and evaluate the adoption of Mamba as an alternative fusion mechanism for brain tumor segmentation in the presence of missing modalities. Experimental results demonstrate that transformer-based architectures achieve leading performance on BraTS2023, outperforming purely convolutional models that were instead superior in BraTS2018. Meanwhile, the proposed Mamba-based architecture exhibits promising performance in comparison to state-of-the-art models, competing and even outperforming transformers. The source code of the proposed approach is publicly released alongside the benchmark developed for the evaluation: https://github.com/AImageLab-zip/IM-Fuse.

The segmentation of the Inferior Alveolar Canal (IAC) plays a central role in maxillofacial surgery, drawing significant attention in the current research. Because of their outstanding results, deep learning methods are widely adopted in the segmentation of 3D medical volumes, including the IAC in Cone Beam Computed Tomography (CBCT) data. One of the main challenges when segmenting large volumes, including those obtained through CBCT scans, arises from the use of patch-based techniques, mandatory to fit memory constraints. Such training approaches compromise neural network performance due to a reduction in the global contextual information. Performance degradation is prominently evident when the target objects are small with respect to the background, as it happens with the inferior alveolar nerve that develops across the mandible, but involves only a few voxels of the entire scan. In order to target this issue and push state-of-the-art performance in the segmentation of the IAC, we propose an innovative approach that exploits spatial information of extracted patches and integrates it into a Transformer architecture. By incorporating prior knowledge about patch location, our model improves state of the art by ~2 points on the Dice score when integrated with the standard U-Net architecture. The source code of our proposal is publicly released.

Recently, Multimodal Large Language Models (MLLMs) have emerged as a leading framework for enhancing the ability of Large Language Models (LLMs) to interpret non-linguistic modalities. Despite their impressive capabilities, the robustness of MLLMs under conditions where one or more modalities are missing remains largely unexplored. In this paper, we investigate the extent to which MLLMs can maintain performance when faced with missing modality inputs. Moreover, we propose a novel framework to mitigate the aforementioned issue called Retrieval-Augmented Generation for missing modalities (MissRAG). It consists of a novel multimodal RAG technique alongside a tailored prompt engineering strategy designed to enhance model robustness by mitigating the impact of absent modalities while preventing the burden of additional instruction tuning. To demonstrate the effectiveness of our techniques, we conducted comprehensive evaluations across five diverse datasets, covering tasks such as audio-visual question answering, audio-visual captioning, and multimodal sentiment analysis.

Cone-beam computed tomography (CBCT) is a standard imaging modality in orofacial and dental practices, providing essential 3D volumetric imaging of anatomical structures, including jawbones, teeth, sinuses, and neurovascular canals. Accurately segmenting these structures is fundamental to numerous clinical applications, such as surgical planning and implant placement. However, manual segmentation of CBCT scans is time-intensive and requires expert input, creating a demand for automated solutions through deep learning. Effective development of such algorithms relies on access to large, well-annotated datasets, yet current datasets are often privately stored or limited in scope and considered structures, especially concerning 3D annotations. This paper proposes ToothFairy2, a comprehensive, publicly accessible CBCT dataset with voxel-level 3D annotations of 42 distinct classes corresponding to maxillofacial structures. We validate the dataset by benchmarking state-of-the-art neural network models, including convolutional, transformer-based, and hybrid Mamba-based architectures, to evaluate segmentation performance across complex anatomical regions. Our work also explores adaptations to the nnU-Net framework to optimize multi-class segmentation for maxillofacial anatomy. The proposed dataset provides a fundamental resource for advancing maxillofacial segmentation and supports future research in automated 3D image analysis in digital dentistry.

This paper tackles the domain of multimodal prompting for visual recognition, specifically when dealing with missing modalities through multimodal Transformers. It presents two main contributions: (i) we introduce a novel prompt learning module which is designed to produce sample-specific prompts and (ii) we show that modality-agnostic prompts can effectively adjust to diverse missing modality scenarios. Our model, termed SCP, exploits the semantic representation of available modalities to query a learnable memory bank, which allows the generation of prompts based on the semantics of the input. Notably, SCP distinguishes itself from existing methodologies for its capacity of self-adjusting to both the missing modality scenario and the semantic context of the input, without prior knowledge about the specific missing modality and the number of modalities. Through extensive experiments, we show the effectiveness of the proposed prompt learning framework and demonstrate enhanced performance and robustness across a spectrum of missing modality cases.

Recently, the field of 3D medical segmentation has been dominated by deep learning models employing Convolutional Neural Networks (CNNs) and Transformer-based architectures, each with its distinctive strengths and limitations. CNNs are constrained by a local receptive field, whereas Transformer are hindered by their substantial memory requirements as well as their data hunger, making them not ideal for processing 3D medical volumes at a fine-grained level. For these reasons, fully convolutional neural networks, as nnU-Net, still dominate the scene when segmenting medical structures in large 3D medical volumes. Despite numerous advancements toward developing transformer variants with subquadratic time and memory complexity, these models still fall short in content-based reasoning. A recent breakthrough is Mamba, a Recurrent Neural Network (RNN) based on State Space Models (SSMs), outperforming Transformers in many long-context tasks (million-length sequences) on famous natural language processing and genomic benchmarks while keeping a linear complexity. In this paper, we evaluate the effectiveness of Mamba-based architectures in comparison to state-of-the-art convolutional and Transformer-based models for 3D medical image segmentation across three well-established datasets: Synapse Abdomen, MSD BrainTumor, and ACDC. Additionally, we address the primary limitations of existing Mamba-based architectures by proposing alternative architectural designs, hence improving segmentation performances. The source code is publicly available to ensure reproducibility and facilitate further research: https://github.com/LucaLumetti/TamingMambas.

Multimodal Large Language Models (MLLMs) have recently emerged as a powerful framework for extending the capabilities of Large Language Models (LLMs) to reason over non-textual modalities. However, despite their success, understanding how they integrate visual and textual information remains an open challenge. Among them, LLaMA~3.2-Vision represents a significant milestone in the development of open-source MLLMs, offering a reproducible and efficient architecture that competes with leading proprietary models, such as Claude 3 Haiku and GPT-4o mini. Motivated by these characteristics, we conduct the first systematic analysis of the information flow between vision and language in LLaMA~3.2-Vision. We analyze three visual question answering (VQA) benchmarks, covering the tasks of VQA on natural images---using both open-ended and multiple-choice question formats---as well as document VQA. These tasks require diverse reasoning capabilities, making them well-suited to reveal distinct patterns in multimodal reasoning. Our analysis unveils a four-stage reasoning strategy: an initial semantic interpretation of the question, an early-to-mid-layer multimodal fusion, a task-specific reasoning stage guided by the resulting multimodal embedding, and a final answer prediction stage. Furthermore, we reveal that multimodal fusion is task-dependent: in complex settings such as document VQA, the model postpones cross-modal integration until semantic reasoning over the question has been established. Overall, our findings offer new insights into the internal dynamics of MLLMs and contribute to advancing the interpretability of vision-language architectures. Our source code is available at https://github.com/AImageLab/MLLMs-FlowTracker.