With over 85 million CT scans performed annually in the United States, creating tumor-related reports is a challenging and time-consuming task for radiologists. To address this need, we present RadGPT, an Anatomy-Aware Vision-Language AI Agent for generating detailed reports from CT scans. RadGPT first segments tumors, including benign cysts and malignant tumors, and their surrounding anatomical structures, then transforms this information into both structured reports and narrative reports. These reports provide tumor size, shape, location, attenuation, volume, and interactions with surrounding blood vessels and organs. Extensive evaluation on unseen hospitals shows that RadGPT can produce accurate reports, with high sensitivity/specificity for small tumor (<2 cm) detection: 80/73% for liver tumors, 92/78% for kidney tumors, and 77/77% for pancreatic tumors. For large tumors, sensitivity ranges from 89% to 97%. The results significantly surpass the state-of-the-art in abdominal CT report generation. RadGPT generated reports for 17 public datasets. Through radiologist review and refinement, we have ensured the reports' accuracy, and created the first publicly available image-text 3D medical dataset, comprising over 1.8 million text tokens and 2.7 million images from 9,262 CT scans, including 2,947 tumor scans/reports of 8,562 tumor instances. Our reports can: (1) localize tumors in eight liver sub-segments and three pancreatic sub-segments annotated per-voxel; (2) determine pancreatic tumor stage (T1-T4) in 260 reports; and (3) present individual analyses of multiple tumors--rare in human-made reports. Importantly, 948 of the reports are for early-stage tumors.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks---which, differing from algorithms, are more flexible and can support different algorithms—including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

In recent years, AI research has demonstrated enormous potential for the benefit of humanity and society. While often better than its human counterparts in classification and pattern recognition tasks, however, AI still struggles with complex tasks that require commonsense reasoning such as natural language understanding. In this context, the key limitations of current AI models are: dependency, reproducibility, trustworthiness, interpretability, and explainability. In this work, we propose a commonsense-based neurosymbolic framework that aims to overcome these issues in the context of sentiment analysis. In particular, we employ unsupervised and reproducible subsymbolic techniques such as auto-regressive language models and kernel methods to build trustworthy symbolic representations that convert natural language to a sort of protolanguage and, hence, extract polarity from text in a completely interpretable and explainable manner.

We introduce an improved version of the principal path method, an algorithm conceived to find smooth paths between objects in space. Some key steps of the algorithm have been changed, making the solution intrinsically local and preventing it from being attracted by a global manifold. Judiciously performing the initialization step with the Dijkstra algorithm and a proper metric, the functional now only performs a final refinement of the initial solution. Hence the algorithm is stabler as the space of possible solutions has been considerably reduced with respect to the original method. We tested the proposed algorithm in 2D toy data sets (to understand the behaviour) and in high-dimensional data sets. Compared to the previous version of the algorithm, we obtained significantly stabler and more realistic generated samples.

Feature selection is a prominent activity when dealing with classification/regression problems in biological and omics data. Despite the effort devoted to this issue theoretically, feature selection stability within and across methods is often overlooked in practice. This is a compelling issue because a unique or at least stable answer is needed in clinical scenarios. Here, we analyse in detail a multiomics small sample data set, the Oxford Street II data set, and discuss how existing methods perform in terms of the usual metrics but also in terms of intra and inter feature selection stability. To mitigate the observed instability, we propose a simple unsupervised feature prefiltering, achieving promising results.

SUMMARY: A primary problem in high-throughput genomics experiments is finding the most important genes involved in biological processes (e.g. tumor progression). In this applications note, we introduce spathial, an R package for navigating high-dimensional data spaces. spathial implements the Principal Path algorithm, which is a topological method for locally navigating on the data manifold. The package, together with the core algorithm, provides several high-level functions for interpreting the results. One of the analyses we propose is the extraction of the genes that are mainly involved in the progress from one state to another. We show a possible application in the context of tumor progression using RNA-Seq and single-cell datasets, and we compare our results with two commonly used algorithms, edgeR and monocle3, respectively. AVAILABILITY AND IMPLEMENTATION: The R package spathial is available on the Comprehensive R Archive Network (https://cran.r-project.org/web/packages/spathial/index.html) and on GitHub (https://github.com/erikagardini/spathial). It is distributed under the GNU General Public License (version 3). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: NanoShaper is a program specifically aiming the construction and analysis of the molecular surface of nanoscopic systems. It uses ray-casting for parallelism and it performs analytical computations whenever possible to maximize robustness and accuracy of the approach. Among the other features, NanoShaper provides volume, surface area, including that of internal cavities, for any considered molecular system. It identifies pockets via a very intuitive definition based on the concept of probe radius, intrinsic to the definition of the solvent excluded surface. We show here that, with a suitable choice of the parameters, the same approach can also permit the visualisation of molecular channels. NanoShaper has now been interfaced with the widely used molecular visualization software VMD, further enriching its already well furnished toolset. AVAILABILITY AND IMPLEMENTATION: VMD is available at http://www.ks.uiuc.edu/Research/vmd/. NanoShaper, its documentation, tutorials and supporting programs are available at http://concept.iit.it/downloads. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

In this paper, we introduce the concept of principal paths in data space; we show that this is a well-characterized problem from the point of view of cognition, and that it can lead to salient insights in the analyzed data enabling topological/holistic descriptions. These paths, interestingly, can be interpreted as local principal curves, and in this paper, we suggest that they are analogous to what, in the statistical mechanics realm, are called minimum free-energy paths. Here, we move that concept from physics to data space and compute them in both the original and the kernel space. The algorithm is a regularized version of the well-known k -means clustering algorithm. The regularization parameter is derived via an in-sample model selection process based on the Bayesian evidence maximization. Interestingly, we show that this choice for the regularization parameter consistently leads to the same manifold even when changing the number of clusters. We apply the method to common data sets, dynamical systems, and, in particular, to molecular dynamics trajectories showing the generality, the usefulness of the approach and its superiority with respect to other related approaches.

Recognizing the samples belonging to one class in a heterogeneous data set is a very interesting but tough machine learning task. Some samples of the data set can be actual outliers or members of other classes for which training examples are lacking. In contrast to other kernel approaches present in the literature, in this work, the problem is faced defining a one-class kernel machine that delivers the probability for a sample to belong to the support of the distribution and that can be efficiently trained by a hybrid sequential minimal optimization-expectation maximization algorithm. Due to the analogy to the import vector machine and to the one-class approach, we named the method import vector domain description (IVDD). IVDD was tested on a toy 2-D data set in order to characterize its behavior on a set of widely used benchmarking UCI data sets and, lastly, challenged against a real world outlier detection data set. All the results were compared against state-of-the-art closely related methods such as one-class-SVM and Support Vector Domain Description, proving that the algorithm is equally accurate with the additional advantage of delivering the probability estimate for each sample. Finally, a few variants aimed at providing memory savings and/or computational speed-up in the light of big data analysis are briefly sketched.

Inductive bias represents an important factor in learning theory, as it can shape the generalization properties of a learning machine. This paper shows that biased regularization can be used as inductive bias to effectively tackle the semi-supervised classification problem. Thus, semi-supervised learning is formalized as a supervised learning problem biased by an unsupervised reference solution. The proposed framework has been tested on a malware-detection problem. Experimental results confirmed the effectiveness of the semi-supervised methodology presented in this paper.

SUMMARY A proper and efficient representation of possibly complex and large molecular surfaces is an important task in bioinformatics and biophysics. Molecular surfaces indeed are used for different aims, in particular, for computation, as visual support tools for biologists, in electrostatics problems involving implicit solvents (e.g. while solving the Poisson–Boltzmann equation) or for molecular dynamics simulations. This is the reason why, in literature, a multitude of algorithms that differ on the basis of the adopted representation and the approach/technology used were proposed. In this paper we present a CUDA‐based software component able to produce high‐resolution molecular surfaces based on the Van der Waals, solvent accessible, Richards–Connolly and blobby definitions. The component was designed to be used in heterogeneous visualization pipelines; therefore, the representation of the molecular surfaces suits both the direct visualization and the efficient storing for following processing steps. Experimental results show speedup figures between 39.3 and 88.4 considering molecules of different sizes and surface resolutions, resulting in meshes of up to 258.6 million triangles. Copyright © 2013 John Wiley & Sons, Ltd.

In this work an implementation of a linearized Poisson-Boltzmann equation solver based on a Finite Differences scheme on the GPU architecture is presented. The algorithm exploits the checkerboard structure of the discretized Laplace operator and follows the footprints of a popular solver called DelPhi, which is widely used in the Computational Biology community. The algorithm has been implemented using CUDA. This implementation has then been integrated with the DelPhi solver and tested over a few representative cases of biological interest. Details of the implementation as well as performance test results are illustrated.

Magnetic-based detection technologies for undersea protection systems are very effective in monitoring critical areas where weak signal sources are difficult to identify (e.g. diver intrusion in proximity of the seafloor). The complexity of the involved geomagnetic phenomena and the nature of the target detection strategy require the use of adaptive methods for signal processing. The paper shows that Computational Intelligence (CI) models can be integrated with those magnetic-based technologies, and presents an effective, reliable system for adaptive undersea protection. Two different CI paradigms are successfully tested for the specific application task: Circular BackPropagation (CBP) and Support Vector Machines (SVMs). Experimental results on real data prove the advantage of the integrated approach over existing conventional methods. Individual CI components and the overall detection system have been verified in real experiments.

Data-intensive applications use empirical methods to extract consistent information from huge samples. When applied to classification tasks, their aim is to optimize accuracy on unseen data hence a reliable prediction of the generalization error is of paramount importance. Theoretical models, such as Statistical Learning Theory, and empirical estimations, such as cross-validation, can both fit data-mining classification domains very well, provided some crucial assumptions are verified in advance. In particular, the stationary distribution of the observed data is critical, although it is sometimes overlooked in practice. The paper formulates an operative criterion to verify the stationary assumption; the method applies to both theoretical and practical predictions of generalization errors. The analysis addresses the specific case of clustering-based classifiers; the K-Winner Machine (KWM) model is used as a reference for its known theoretical bounds; cross-validation provides an empirical counterpart for practical comparison. The criterion, based on efficient unsupervised clustering-based probability distribution estimation, is tested experimentally on a set of different, data-intensive applications, including: intrusion detection for computer-network security, optical character recognition, text mining and pedestrian detection. Experimental results confirm the effectiveness of the proposed approach to efficiently detect non stationarity.

The two major components of a robotic tactile-sensing system are the tactile-sensing hardware at the lower level and the computational/software tools at the higher level. Focusing on the latter, this research assesses the suitability of computational-intelligence (CI) tools for tactile-data processing. In this context, this paper addresses the classification of sensed object material from the recorded raw tactile data. For this purpose, three CI paradigms, namely, the support-vector machine (SVM), regularized least square (RLS), and regularized extreme learning machine (RELM), have been employed, and their performance is compared for the said task. The comparative analysis shows that SVM provides the best tradeoff between classification accuracy and computational complexity of the classification algorithm. Experimental results indicate that the CI tools are effective in dealing with the challenging problem of material classification.

Neural Networks are powerful tools for function approximation problems. A possible peculiar application of neural networks is that proposed here: estimating the univariate mean of a distribution from a finite sample. This problem characterizes a huge number of applicative and scientific problems. The Gaussian distribution case is analyzed, however the proposed analysis is of general validity and can be easily extended to other distributions. In particular the estimation problem is approached as a regularization problem and a solution to the selection of the regularization parameter is obtained via the employment of neural models. The paper, after introducing some theoretical results, presents two neural models, namely a MLP and a Circular Back Propagation Network, for the mean prediction. Experimental results show that neural networks can estimate the mean, in expectation, better than the usual sample mean formula.

A crucial issue in designing learning machines is to select the correct model parameters. When the number of available samples is small, theoretical sample-based generalization bounds can prove effective, provided that they are tight and track the validation error correctly. The maximal discrepancy (MD) approach is a very promising technique for model selection for support vector machines (SVM), and estimates a classifier's generalization performance by multiple training cycles on random labeled data. This paper presents a general method to compute the generalization bounds for SVMs, which is based on referring the SVM parameters to an unsupervised solution, and shows that such an approach yields tight bounds and attains effective model selection. When one estimates the generalization error, one uses an unsupervised reference to constrain the complexity of the learning machine, thereby possibly decreasing sharply the number of admissible hypothesis. Although the methodology has a general value, the method described in the paper adopts vector quantization (VQ) as a representation paradigm, and introduces a biased regularization approach in bound computation and learning. Experimental results validate the proposed method on complex real-world data sets.

A central issue in computational intelligence is the training phase of a learning machine. In classification problems, in particular, Support Vector Machines are one of the most effective tools. In this work an analog low-complexity circuital implementation is proposed to address the learning stage of SVMs. The circuit is a co-content minimization network based on a suitable SVM formulation embedding bias removal. Moreover the circuit complexity (i.e. the density of the kernel matrix) is effectively controlled by resorting to a proper kernel function. Experimental evidence shows the effectiveness of the proposed approach.

Regularized classifiers such as SVM or RLS are among the most used and successful classifiers in machine learning. The theory and the empirical evaluation of the associate generalization bounds are of paramount importance; bounds based on the Maximal-Discrepancy approach proved quite effective. The paper shows an efficient, iterative procedure to evaluate Maximal-Discrepancy bounds for this kind of classifiers. Empirical results on UCI datasets show that this approach can attain tighter bounds to the run-time classification error.

Training Support Vector Machines (SVMs) requires efficient architectures, endowed with agile memory handling and specific computational features. Such a process is often supported by embedded implementations on dedicated machinery, for example in applications requiring on-line training abilities. The paper presents a general approach to the efficient implementation of SVM training on Digital Signal Processor (DSP) devices. The methodology optimizes efficiency by a twofold approach: first, it suitably adjusts an established, effective training algorithm for large data sets; secondly, it reformulates the algorithm to best exploit the computational features of DSP devices and boost efficiency accordingly. Experimental results tackle the training problem of SVMs by using a high-end DSP architecture on real-world benchmarks, and confirm both the effectiveness and the general validity of the approach.