Search

ABSTRACT: Background: In recent years, the integration of Artificial Intelligence (AI) models has revolutionized the diagnosis of Low Back Pain (LBP) and associated disc pathologies. Among these, SpineNetV2 stands out as a state-of-the-art, open-access model for detecting and grading various intervertebral disc pathologies. However, ensuring the reliability and applicability of AI models like SpineNetV2 is paramount. Rigorous validation is essential to guarantee their robustness and generalizability across diverse patient cohorts and imaging protocols. Methods: We conducted a retrospective analysis of MRI images of 1747 lumbosacral intervertebral discs (IVDs) from 353 patients (mean age, 54 ± 15.4 years, 44.5% female) with various spinal disorders, collected between September 2021 and February 2023 at X-Ray Service s.r.l. The SpineNetV2 system was used to grade 11 distinct lumbosacral disc pathologies, including Pfirrmann grading, disc narrowing, central canal stenosis, spondylolisthesis, (upper and lower) endplate defects, (upper and lower) marrow changes, (right and left) foraminal stenosis, and disc herniation, using T2-weighted sagittal MR images. Performance metrics included accuracy, balanced accuracy, precision, F1 score, Matthew's Correlation Coefficient, Brier Score Loss, Lin's concordance correlation coefficients, and Cohen's kappa coefficients. Two expert radiologists provide annotations for these discs. The evaluation of SpineNetV2′s grading is compared against expert radiologists' assessments. Results: SpineNetV2 demonstrated strong performance across various metrics, with high agreement scores (Cohen's Kappa, Lin's Concordance, and Matthew's Correlation Coefficient exceeding 0.7) for most pathologies. However, lower agreement was found for foraminal stenosis and disc herniation, underscoring the limitations of sagittal MR images for evaluating these conditions. Conclusions: This study highlights the importance of external validation, emphasizing the need for comprehensive assessments of deep learning models. SpineNetV2 exhibits promising results in predicting disc pathologies, with findings guiding further improvements. The open-source release of SpineNetV2 enables researchers to independently validate and extend the model's capabilities. This collaborative approach promotes innovation and accelerates the development of more reliable and comprehensive deep learning tools for the assessment of spine pathology.

In this work, we present a novel dataset, SynthOutdoor, comprising 39,086 high-resolution images, aimed at addressing the data scarcity in the field of 3D light direction estimation under the assumption of distant lighting. SynthOutdoor was generated using our software (which is also publicly available), that in turn is based on the Unity3D engine. Our dataset provides a set of images rendered from a given input scene, with the camera moving across a predefined path within the scene. This dataset captures a wide variety of lighting conditions through the implementation of a solar cycle.The dataset's ground truth is composed of the following elements: the 3D light direction and color intensity of the sun; the color intensity of the ambient light; the instance segmentation masks of each object and the surface normals map, in which each pixel is assigned with the 3D surface normal in that point (encoded as 3 color channels).By providing not only the light direction and intensity, but also the geometric and semantic information of the rendered images, our dataset can be used not only for light estimation, but also for more general tasks such as 3D geometry and shading estimation from 2D images.

In this paper, a framework for image and video intra-frame coding able to effectively employ the multiple transform paradigm using Symmetry-Based Graph Fourier Transforms (SBGFTs) is proposed. As data representation relies heavily on the characteristics of signal classes in high dimensional spaces, over the years it has been understood that signal instances for a given class, e.g., images, typically lie on a manifold which is not a single linear subspace. Accordingly, standards for image/video compression have considered to introduce multiple representation models to encode the data, i.e., multiple linear block transforms. However, the advantages are currently limited typically due to implementation complexity and the high signaling cost of the representation mode. As a result, only a small set of alternative transforms is typically considered, which restricts the adaptation capabilities to the data. In this paper, we instead argue that it is feasible to incorporate into the image coding framework a large set of SBGFTs which can overcome the aforementioned drawbacks. Specifically, we demonstrate the ability to derive the block encoding transform index from the quantization pattern fingerprint using a Multilayer Perceptron (MLP) architecture, achieving prediction with high confidence, surpassing 80% for 4-top accuracy. This translates into a more efficient entropy coding strategy for the index, allowing to save on the average more than 3 bits per block. These experimental results highlight the significant benefits of SBGFTs for multiple transform coding, particularly when combined with the proposed MLP based index prediction, with only a limited increase in computational complexity.

Abstract Ultrasonic vocalizations (USVs) analysis represents a fundamental tool to study animal communication. It can be used to perform a behavioral investigation of mice for ethological studies and in the field of neuroscience and neuropharmacology. The USVs are usually recorded with a microphone sensitive to ultrasound frequencies and then processed by specific software, which help the operator to identify and characterize different families of calls. Recently, many automated systems have been proposed for automatically performing both the detection and the classification of the USVs. Of course, the USV segmentation represents the crucial step for the general framework, since the quality of the call processing strictly depends on how accurately the call itself has been previously detected. In this paper, we investigate the performance of three supervised deep learning methods for automated USV segmentation: an Auto-Encoder Neural Network (AE), a U-NET Neural Network (UNET) and a Recurrent Neural Network (RNN). The proposed models receive as input the spectrogram associated with the recorded audio track and return as output the regions in which the USV calls have been detected. To evaluate the performance of the models, we have built a dataset by recording several audio tracks and manually segmenting the corresponding USV spectrograms generated with the Avisoft software, producing in this way the ground-truth (GT) used for training. All three proposed architectures demonstrated precision and recall scores exceeding $$90\%$$ 90 % , with UNET and AE achieving values above $$95\%$$ 95 % , surpassing other state-of-the-art methods that were considered for comparison in this study. Additionally, the evaluation was extended to an external dataset, where UNET once again exhibited the highest performance. We suggest that our experimental results may represent a valuable benchmark for future works.

Along the Mediterranean Sea shelf, algal reefs made of crustose coralline algae and Peyssonneliales are known as Coralligenous. It ranks among the most important ecosystems in the Mediterranean Sea because of its extent, complexity, and heterogeneity, supporting very high levels of biodiversity. Descriptive approaches for monitoring purposes are often aimed at assessing the surficial ephemeral canopy, which is sustained and controlled by the occurrence of the long-lasting rigid structure at the base. This practice led to the non-univocal definition of Coralligenous, sometimes indicated as “animal Coralligenous” because of the surficial dominance of these components. The quantitative assessment of the builders that actively build up the persistent structure through geological time is therefore a fundamental topic. We collected two discrete coralligenous samples in front of Marzamemi village (Sicily, Ionian Sea), the first from an area of a dense coralligenous cover (- 37 m) and the second one from an area with sparse build-ups (- 36 m). By using image analysis and computerized axial tomography, we distinguished and quantified the different components both on the surface and inside the framework. In both cases, our results confirm the primary role of crustose coralline algae as major builders of the Mediterranean Coralligenous, this aspect matching with the evidence from the Quaternary fossil record. We suggest that the role of encrusting calcareous red algae in the Coralligenous should be considered in conservation and management policies.

Ultrasonic vocalizations (USVs) analysis is a well-recognized tool to investigate animal communication. It can be used for behavioral phenotyping of murine models of different disorders. The USVs are usually recorded with a microphone sensitive to ultrasound frequencies and they are analyzed by specific software. Different calls typologies exist, and each ultrasonic call can be manually classified, but the qualitative analysis is highly time-consuming. Considering this framework, in this work we proposed and evaluated a set of supervised learning methods for automatic USVs classification. This could represent a sustainable procedure to deeply analyze the ultrasonic communication, other than a standardized analysis. We used manually built datasets obtained by segmenting the USVs audio tracks analyzed with the Avisoft software, and then by labelling each of them into 10 representative classes. For the automatic classification task, we designed a Convolutional Neural Network that was trained receiving as input the spectrogram images associated to the segmented audio files. In addition, we also tested some other supervised learning algorithms, such as Support Vector Machine, Random Forest and Multilayer Perceptrons, exploiting informative numerical features extracted from the spectrograms. The performance showed how considering the whole time/frequency information of the spectrogram leads to significantly higher performance than considering a subset of numerical features. In the authors' opinion, the experimental results may represent a valuable benchmark for future work in this research field.

The FASSEG repository is composed by four subsets containing face images useful for training and testing automatic methods for the task of face segmentation. Threesubsets, namely frontal01, frontal02, and frontal03 are specifically built for performing frontal face segmentation. Frontal01 contains 70 original RGB images and the corresponding roughly labelledground-truth masks. Frontal02 contains the same image data, with high-precision labelled ground-truth masks. Frontal03 consists in 150 annotated face masks of twins captured in various orientations, illumination conditions and facial expressions. The last subset, namely multipose01, contains more than 200 faces in multiple poses and the corresponding ground-truth masks. For all face images, ground-truth masks are labelled on six classes (mouth, nose, eyes, hair, skin, and background).

We present a statistical framework based on Hidden Markov Models (HMMs) for skimming feature films. A chain of HMMs is used to model subsequent story units: HMM states represent different visual-concepts, transitions model the temporal dependencies in each story unit, and stochastic observations are given by single shots. The skim is generated as an observation sequence, where, in order to privilege more informative segments for entering the skim, shots are assigned higher probability of observation if endowed with salient features related to specific film genres. The effectiveness of the method is demonstrated by skimming the first thirty minutes of a wide set of action and dramatic movies, in order to create previews for users useful for assessing whether they would like to see that movie or not, but without revealing the movie central part and plot details. Results are evaluated and compared through extensive user tests in terms of metrics that estimate the content representational value of the obtained video skims and their utility for assessing the user's interest in the observed movie.

In distributed video coding the signal prediction is shifted at the decoder side, giving therefore most of the computational complexity burden at the receiver. Moreover, since no prediction loop exists before transmission, an intrinsic robustness to transmission errors has been claimed. This work evaluates and compares the error resilience performance of two distributed video coding architectures. In particular, we have considered a video codec based on the Stanford architecture (DISCOVER codec) and a video codec based on the PRISM architecture. Specifically, an accurate temporal and rate/distortion based evaluation of the effects of the transmission errors for both the considered DVC architectures has been performed and discussed. These approaches have been also compared with H.264/AVC, in both cases of no error protection, and simple FEC error protection. Our evaluations have highlighted in all cases a strong dependence of the behavior of the various codecs to the content of the considered video sequence. In particular, PRISM seems to be particularly well suited for low-motion sequences, whereas DISCOVER provides better performance in the other cases.

Indexing deals with the automatic extraction of information with the objective of automatically describing and organizing the content. Thinking of a video stream, different types of information can be considered semantically important. Since we can assume that the most relevant one is linked to the presence of moving foreground objects, their number, their shape, and their appearance can constitute a good mean for content description. For this reason, we propose to combine both motion information and region-based color segmentation to extract moving objects from an MPEG2 compressed video stream starting only considering low-resolution data. This approach, which we refer to as “rough indexing,†consists in processing P-frame motion information first, and then in performing I-frame color segmentation. Next, since many details can be lost due to the low-resolution data, to improve the object detection results, a novel spatiotemporal filtering has been developed which is constituted by a quadric surface modeling the object trace along time. This method enables to effectively correct possible former detection errors without heavily increasing the computational effort.