Accurate 3D segmentation of multiple sclerosis lesions is critical for clinical practice, yet existing approaches face key limitations: many models rely on 2D architectures or partial modality combinations, while others struggle to generalise across scanners and protocols. Although large-scale, multi-site training can improve robustness, its data demands are often prohibitive. To address these challenges, we propose a 3D multi-modal network that simultaneously processes T1-weighted, T2-weighted, and FLAIR scans, leveraging full cross-modal interactions and volumetric context to achieve state-of-the-art performance across four diverse public datasets. To tackle data scarcity, we quantify the minimal fine-tuning effort needed to adapt to individual unseen datasets and reformulate the few-shot learning paradigm at an “instance-per-dataset” level (rather than traditional “instance-per-class”), enabling the quantification of the minimal fine-tuning effort to adapt to multiple unseen sources simultaneously. Finally, we introduce Latent Distance Analysis , a novel label-free reliability estimation technique that anticipates potential distribution shifts and supports any form of test-time adaptation, thereby strengthening efficient robustness and physicians’ trust.

We introduce a novel deep neural network designed for fast and structurally consistent monocular 360° depth estimation in indoor settings. Our model generates a spherical depth map from a single gravity-aligned or gravity-rectified equirectangular image, ensuring the predicted depth aligns with the typical depth distribution and structural features of cluttered indoor spaces, which are generally enclosed by walls, floors, and ceilings. By leveraging the distinctive vertical and horizontal patterns found in man-made indoor environments, we propose a streamlined network architecture that incorporates gravity-aligned feature flattening and specialized vision transformers. Through flattening, these transformers fully exploit the omnidirectional nature of the input without requiring patch segmentation or positional encoding. To further enhance structural consistency, we introduce a novel loss function that assesses density map consistency by projecting points from the predicted depth map onto a horizontal plane and a cylindrical proxy. This lightweight architecture requires fewer tunable parameters and computational resources than competing methods. Our comparative evaluation shows that our approach improves depth estimation accuracy while ensuring greater structural consistency compared to existing methods. For these reasons, it promises to be suitable for incorporation in real-time solutions, as well as a building block in more complex structural analysis and segmentation methods.

Many clinical and research studies of the human brain require accurate structural MRI segmentation. While traditional atlas-based methods can be applied to volumes from any acquisition site, recent deep learning algorithms ensure high accuracy only when tested on data from the same sites exploited in training (i.e., internal data). Performance degradation experienced on external data (i.e., unseen volumes from unseen sites) is due to the inter-site variability in intensity distributions, and to unique artefacts caused by different MR scanner models and acquisition parameters. To mitigate this site-dependency, often referred to as the scanner effect, we propose LOD-Brain, a 3D convolutional neural network with progressive levels-of-detail (LOD), able to segment brain data from any site. Coarser network levels are responsible for learning a robust anatomical prior helpful in identifying brain structures and their locations, while finer levels refine the model to handle site-specific intensity distributions and anatomical variations. We ensure robustness across sites by training the model on an unprecedentedly rich dataset aggregating data from open repositories: almost 27,000 T1w volumes from around 160 acquisition sites, at 1.5 - 3T, from a population spanning from 8 to 90 years old. Extensive tests demonstrate that LOD-Brain produces state-of-the-art results, with no significant difference in performance between internal and external sites, and robust to challenging anatomical variations. Its portability paves the way for large-scale applications across different healthcare institutions, patient populations, and imaging technology manufacturers. Code, model, and demo are available on the project website.

The perceived distance of the camera from the subject of a filmed scene, namelyshot scale, is a prominent formal feature of any filmic product, endowed with both stylistic and narrative functions. To measure how shot scale affects both lower and higher complexity responses in viewers, we first investigate how the distribution and rotation of Close, Medium, and Long Shots relate to viewers’ rating onfilm mood, assessed in terms of hedonic tone, energetic arousal, and tense arousal on an extensive set of 50 film clips. Then we examine the effect of shot scale on viewers of violent scenes in terms ofnarrative engagementand its sub-scales: narrative understanding, attentional focus, emotional engagement, and narrative presence. To enable such analysis, shot scale classification is automatized by means of Convolutional Neural Networks trained on the filmographies by six directors for a total of 120 full-length movies analysed at one frame per second. Based on large corpora, this study provides methods for further investigating the relationship between shot scale and the viewers’ emotional involvement. Beyond stylistic analysis, gaining valuable insight into the mechanism of narrative effects of film stories could be useful for purposes such as movie recommendation and film therapy.

Many functional and structural neuroimaging studies call for accurate morphometric segmentation of different brain structures starting from image intensity values of MRI scans. Current automatic (multi-) atlas-based segmentation strategies often lack accuracy on difficult-to-segment brain structures and, since these methods rely on atlas-to-scan alignment, they may take long processing times. Alternatively, recent methods deploying solutions based on Convolutional Neural Networks (CNNs) are enabling the direct analysis of out-of-the-scanner data. However, current CNN-based solutions partition the test volume into 2D or 3D patches, which are processed independently. This process entails a loss of global contextual information, thereby negatively impacting the segmentation accuracy. In this work, we design and test an optimised end-to-end CNN architecture that makes the exploitation of global spatial information computationally tractable, allowing to process a whole MRI volume at once. We adopt a weakly supervised learning strategy by exploiting a large dataset composed of 947 out-of-the-scanner (3 Tesla T1-weighted 1mm isotropic MP-RAGE 3D sequences) MR Images. The resulting model is able to produce accurate multi-structure segmentation results in only a few seconds. Different quantitative measures demonstrate an improved accuracy of our solution when compared to state-of-the-art techniques. Moreover, through a randomised survey involving expert neuroscientists, we show that subjective judgements favour our solution with respect to widely adopted atlas-based software.

The apparent distance of the camera from the subject of a filmed scene, namely shot scale, is one of the prominent formal features of any filmic product, endowed with both stylistic and narrative functions. In this work we propose to use Convolutional Neural Networks for the automatic classification of shot scale into Close-, Medium-, or Long-shots. The development of such a tool allows for investigating the relationship between shot scale computed of large movie corpora and the viewers' emotional involvement, for purposes such as movie recommendation, stylistic analysis, and film therapy, to name a few. Training and testing are performed on the full filmographies by six different authors (Scorsese, Godard, Tarr, Fellini, Antonioni, and Bergman) for a total number of 120 movies analysed second by second. Classification results are widely superior to state-of-the-art, with an overall accuracy around 94%.

Huge streams of diagnostic images are expected to be produced daily in the emerging field of digital microbiology imaging because of the ongoing worldwide spread of Full Laboratory Automation systems. This is redefining the way microbiologists execute diagnostic tasks. In this context, the authors want to assess the suitability and effectiveness of a deep learning approach to solve the diagnostically relevant but visually challenging task of directly identifying pathogens on bacterial growing plates. In particular, starting from hyperspectral acquisitions in the VNIR range and spatial‐spectral processing of cultured plates, they approach the identification problem as the classification of computed spectral signatures of the bacterial colonies. In a highly relevant clinical context (urinary tract infections) and on a database of acquired hyperspectral images, they designed and trained a convolutional neural network for pathogen identification, assessing its performance and comparing it against conventional classification solutions. At the same time, given the expected data flow and possible conservation and transmission needs, they are interested in evaluating the combined use of classification and lossy data compression. To this end, after selecting a suitable wavelet‐based compression technology, they test coding strength‐driven operating points looking for configurations able to provably prevent any classification performance degradation.

Working with noisy meshes and aiming at providing high-fidelity 3D object models without tampering the metric quality of the acquisitions, we propose a mesh denoising technique that, through a normal-diffusion process guided by a curvature saliency map, is able to preserve and emphasize the natural object features, concurrently allowing the introduction of a bound on the maximum distance from the original model. Moreover, both the position of the mesh vertices and the edge orientations are optimized through a tailored geometric-aliasing correction. Thanks to an efficiently parallelized procedure, we are able to process even large models almost instantly with a parameter configuration that does not depend on the scale of the object. An essential survey on mesh denoising is also presented which is functional to the definition of a common framework where to set up our solutions and the related technical and experimental comparisons. The proposed results prove the effectiveness of our method, especially on the challenging target application profiles. Where competing techniques tend to inappropriately recover sharp edges while deforming the surrounding geometry or, on the contrary, to oversmooth shallow features, our method protects and enhances the natural object features and effectively reduces scanning noise on the smooth parts, while guaranteeing the prescribed metric-fidelity to the input model.

Multi-View Stereo (MVS) algorithms scale poorly on large image sets, and quickly become unfeasible to run on a single machine with limited memory. Typical solutions to lower the complexity include reducing the redundancy of the image set (view selection), and dividing the image set in groups to be processed independently (view clustering). A novel formulation for view selection is proposed here. We express the problem with an Integer Linear Programming (ILP) model, where cameras are modeled with binary variables, while the linear constraints enforce the completeness of the 3D reconstruction. The solution of the ILP leads to an optimal subset of selected cameras. As a second contribution, we integrate ILP camera selection with a view clustering approach which exploits Leveraged Affinity Propagation (LAP). LAP clustering can efficiently deal with large camera sets. We adapt the original algorithm so that it provides a set of overlapping clusters where the minimum and maximum sizes and the number of overlapping cameras can be specified. Evaluations on four different dataset show our solution provides significant complexity reductions and guarantees near-perfect coverage, making large reconstructions feasible even on a single machine.

In this work we compare state-of-the-art 3D coding technologies which are suitable for compression of 3D (volumetric) medical imagery. Our aim is to assess the usage of these methods on representative large datasets, also keeping into account relevant features related to modern application requirements. This work wants to be a useful reference for those interested in medical image coding technologies evaluation and, more in general, in the interdisciplinary debate about the usage of irreversible compression technologies for an improved and cost effective handling of diagnostic imaging processes and infrastructures (PACS, teleradiology) involving large datasets. Reproducibility and possible extension of this study are guaranteed by the use of publicly available reference software and datasets.

Aiming at reaching an interactive and simplified usage of high-resolution 3D acquisition systems, this paper presents a fast and automated technique for pre-alignment of dense range images. Starting from a multi-scale feature point extraction and description, a processing chain composed by feature matching and correspondence searching, ranking grouping and skimming is performed to select the most reliable correspondences over which the correct alignment is estimated. Pre-alignment is obtained in few seconds per million point images on a off-the-shelf PC architecture. The experimental setup aimed to demonstrate the system behavior with respect to a set of concurrent requirements and the obtained performance are significant in the perspective of a fast, robust and unconstrained 3D object reconstruction.

Modern medical imaging modalities produce increasingly large datasets. This trend can be in contrast with the computation and transmission time requirements coming from critical teleradiology applications. Multidimensional image compression techniques can be considered as enabling solutions on condition that they are able to guarantee a suitable combination of rate-distortion and computational performance which fulfill all the application domain requirements. In this work, we present a parallel version of our 3D Embedded Morphological Dilation Coding algorithm that allows a significant reduction of computation costs and the concurrent conservation of coding performance and of other relevant bitstream properties. A comparison with the recently released JPEG2000 part 10 (JP3D) standard put in evidence the value of the proposed solution, especially for teleradiology applications over heterogeneous networks.

We present an improvement of the classical marker-controlled watershed approach in the direction of a better exploitation of user-defined markers. The combined action of a partial flooding and paint drops falling downwards on the gray value relief from marker locations, leads to a robust and meaningful identification of the candidate basins, which is a prerequisite for an accurate segmentation. This is useful for user-controlled segmentation of biomedical volumes in that it facilitates robust identification of complex 3D structures with inhomogeneous borders. To this end, a visual interactive segmentation system has been implemented where different user-data interaction tools can be selected by physicians to generate machine-understandable knowledge in a quick and compact way. Experimental results on selected use-cases demonstrate the strengths of the proposed solutions.

In the past decades, most of the research on image and video compression has focused on addressing high bandwidth-constrained environments. However, for high resolution and high quality image and video compression, as in the case of high definition television (HDTV) or digital cinema (DC), the primary constraints are related to quality and flexibility. This paper presents a comparison between scalable wavelet-based video codecs and the state of the art in single point encoding and it investigates the obtainable compression efficiency when using temporal correlation with respect to pure intra coding.

Scalable video coding (SVC) differs form traditional single point approaches mainly because it allows to encode in a unique bit stream several working points corresponding to different quality, picture size and frame rate. This work describes the current state-of-the-art in SVC, focusing on wavelet based motion-compensated approaches (WSVC). It reviews individual components that have been designed to address the problem over the years and how such components are typically combined to achieve meaningful WSVC architectures. Coding schemes which mainly differ from the space-time order in which the wavelet transforms operate are here compared, discussing strengths and weaknesses of the resulting implementations. An evaluation of the achievable coding performances is provided considering the reference architectures studied and developed by ISO/MPEG in its exploration on WSVC. The paper also attempts to draw a list of major differences between wavelet based solutions and the SVC standard jointly targeted by ITU and ISO/MPEG. A major emphasis is devoted to a promising WSVC solution, named STP-tool, which presents architectural similarities with respect to the SVC standard. The paper ends drawing some evolution trends for WSVC systems and giving insights on video coding applications which could benefit by a wavelet based approach.

In this letter, an efficient morphological wavelet coder is proposed. The clustering trend of significant coefficients is captured by a new kind of multiresolution binary dilation operator. The layered and adaptive nature of the subband dilation makes it possible for the coding technique to produce an embedded bit-stream with a modest computational cost and state-of-the-art rate-distortion performance. Morphological wavelet coding appears promising because the localized analysis of wavelet coefficient clusters is adequate to capture intrinsic patterns of the source, which can have substantial benefits for reducing further the data redundancy.

We study the effects of the quantization of the three-dimensional wavelet transform coefficients, encoded using a 3D zerotree-based compression scheme. The processing of the quantization error operated by the synthesis filter bank actually determines a modulation of the reconstruction error statistics. This effect entails a non-negligible and potentially objectionable PSNR oscillation among adjacent slices of the decoded 3D data-set, which, e.g. in the biomedical field, would typically be used jointly to reach a diagnosis. In particular, we propose a cyclostationary model of the error statistics fluctuations and an experimental validation of the model itself. Moreover some strategies to reduce such oscillation are suggested and evaluated.

Presents a new method able to convey more meaningful images to the clinician. To reach this goal a new approach to medical interactive segmentation issue is proposed. First an automatic simplification process is performed on the volume, relying on connected operator methods. Afterwards a set of descriptors of the region of interest (RoI) is acquired by means of a 2D interactive session by the clinician. These descriptors, reflecting a real diagnostic interest, work like selective criteria during the subsequent morphological segmentation process done on the whole volume. They allow one to extract the desired structures avoiding a strong interaction by the user. Finally a 3D visualization of the segmented volume of interest (VoI) is obtained, and interactive tools allow the user to adjust and improve the results until satisfaction.

A new way to improve the representation of images using a discrete wavelet transform for coding purposes is presented. The idea lies in combining all wavelet coefficients related to detail information at a same resolution level but along different orientations (horizontal, vertical, and diagonal), into a single image. Given that detail information is located for all subband images in the neighborhood of high frequency textures or edge locations, the pattern of significant coefficients remains unchanged after the combination process. This process allows one to further reduce the number of transformed coefficients by 2/3, while preserving the multiresolution structure. This information can thus be efficiently coded using a multiresolution embedded coding scheme, such as Shapiro's (see IEEE Trans. on Signal Proc., vol.SP-41, no.12, p.3445-62, 1993) zerotree coder. Overall, a higher coding efficiency can be reached while preserving the cross-scale prediction of significance among the coefficients. Ultimately, approximate detail information must be recovered from the combined and coded data for each subband of the original wavelet, so as to reconstruct a decoded image.