In recent years, medical technology companies have increasingly been integrating time-of-flight cameras into their X-ray devices to support optimal collimation and the patient positioning process. However, for many hospitals, it is not financially viable to acquire a new X-ray device with such a camera for those features. In order to still have support for e.g. patient positioning, without having to buy a new X-ray device and be dependent on proprietary algorithms, it is possible to only acquire a time-of-flight camera, attach it to the X-ray device, and use custom algorithms. In this work, we evaluated the ideal camera position for AI-supported patient pose assessment based on depth images for such a setup and assessed the usefulness of a setup with multiple cameras. For this, we generated a total of 461,550 synthetic depth images from CT scans from 50 different camera positions and synthetic radiographs in order to investigate in 1,500 experiments how accurately patients’ poses can be assessed with different camera positions. We found that a camera position perpendicular to the target anatomy being radiographed is particularly well suited, and that adding a second camera does not significantly improve performance.

This paper elucidates that current state-of-the-art Large Language Models (LLMs) are fundamentally incapable of making decisions or developing "thoughts" within the feature space due to their architectural constraints. We establish a definition of "thought" that encompasses traditional understandings of that term and adapt it for application to LLMs. We demonstrate that the architectural design and language modeling training methodology of contemporary LLMs inherently preclude them from engaging in genuine thought processes. Our primary focus is on this theoretical realization rather than practical insights derived from experimental data. Finally, we propose solutions to enable thought processes within the feature space and discuss the broader implications of these architectural modifications.

We study over-parameterized classifiers where empirical risk minimization (ERM) for learning leads to zero training error. In these over-parameterized settings, there are many global minima with zero training error, some of which generalize better than others. We show that under certain conditions, the fraction of “bad” global minima with a true error larger than ε decays to zero exponentially fast with the number of training data n. The bound depends on the distribution of the true error over the set of classifier functions used for the given classification problem, and does not necessarily depend on the size or complexity (e.g., the number of parameters) of the classifier function set. This insight provides an alternative perspective on the unexpectedly good generalization even of highly over-parameterized neural networks. We substantiate our theoretical findings through experiments on synthetic data and a subset of MNIST. Additionally, we assess our hypothesis using VGG19 and ResNet18 on a subset of Caltech101.

This paper proposes a novel regularization approach to bias Convolutional Neural Networks (CNNs) toward utilizing edge and line features in their hidden layers. Rather than learning arbitrary kernels, we constrain the convolution layers to edge and line detection kernels. This intentional bias regularizes the models, improving generalization performance, especially on small datasets. As a result, test accuracies improve by margins of 5-11 percentage points across four challenging fine-grained classification datasets with limited training data and an identical number of trainable parameters. Instead of traditional convolutional layers, we use Pre-defined Filter Modules, which convolve input data using a fixed set of 3x3 pre-defined edge and line filters. A subsequent ReLU erases information that did not trigger any positive response. Next, a 1x1 convolutional layer generates linear combinations. Notably, the pre-defined filters are a fixed component of the architecture, remaining unchanged during the training phase. Our findings reveal that the number of dimensions spanned by the set of pre-defined filters has a low impact on recognition performance. However, the size of the set of filters matters, with nine or more filters providing optimal results.

We evaluate the information that can unintentionally leak into the low dimensional output of a neural network, by reconstructing an input image from a 40- or 32-element feature vector that intends to only describe abstract attributes of a facial portrait. The reconstruction uses blackbox-access to the image encoder which generates the feature vector. Other than previous work, we leverage recent knowledge about image generation and facial similarity, implementing a method that outperforms the current state-of-the-art. Our strategy uses a pretrained StyleGAN and a new loss function that compares the perceptual similarity of portraits by mapping them into the latent space of a FaceNet embedding. Additionally, we present a new technique that fuses the output of an ensemble, to deliberately generate specific aspects of the recreated image.

The correct pose of the patient during radiography is of critical importance to ensure an adequate diagnostic quality of radiographs, which are the basis for diagnosis and treatment planning. However, correct patient positioning is not a standardized process, often resulting in inadequate radiographs and repeated radiation exposure. We propose a novel approach using Time-of-Flight cameras to assess the patient's pose and therefore predict the expected diagnostic quality of the radiograph, before it is even captured. As a first step towards this goal, we acquired a new dataset, consisting of depth images and corresponding radiographs of the ankle using two anatomical preparations in multiple poses. The radiographs were labeled by radiologists for their diagnostic quality related to the patient's pose. These labels serve as quality label for the corresponding pose. Using this dataset we trained deep neural networks and were able to correctly assess the diagnostic quality of a pose with a mean accuracy of up to 90.2%, demonstrating that shared features for pose assessment across patients exist and can be learned.

Sleep spindles are one of the prominent EEG oscillatory rhythms of non-rapid eye movement sleep. In the memory consolidation, these oscillations have an important role in the processes of long-term potentiation and synaptic plasticity. Moreover, the activity (spindle density and/or sigma power) of spindles has a linear association with learning performance in different paradigms. According to the experimental observations, the sleep spindle activity can be improved by closed loop acoustic stimulations (CLAS) which eventually improve memory performance. To examine the effects of CLAS on spindles, we propose a biophysical thalamocortical model for slow oscillations (SOs) and sleep spindles. In addition, closed loop stimulation protocols are applied on a thalamic network. Our model results show that the power of spindles is increased when stimulation cues are applied at the commencing of an SO Down-to-Up-state transition, but that activity gradually decreases when cues are applied with an increased time delay from this SO phase. Conversely, stimulation is not effective when cues are applied during the transition of an Up-to-Down-state. Furthermore, our model suggests that a strong inhibitory input from the reticular (RE) layer to the thalamocortical (TC) layer in the thalamic network shifts leads to an emergence of spindle activity at the Up-to-Down-state transition (rather than at Down-to-Up-state transition), and the spindle frequency is also reduced (8-11 Hz) by thalamic inhibition.

The purpose of this study was to investigate the possibility of implementing an artificial intelligence (AI) approach for the analysis of fluorescence lifetime imaging ophthalmoscopy (FLIO) data even with small data. FLIO data, including the fluorescence intensity and mean fluorescence lifetime (τm) of two spectral channels, as well as OCT-A data from 26 non-smokers and 28 smokers without systemic and ocular diseases were used. The analysis was performed with support vector machines (SVMs), a well-known AI method for small datasets, and compared with the results of convolutional neural networks (CNNs) and autoencoder networks. The SVM was the only tested AI method, which was able to distinguish τm between non-smokers and heavy smokers. The accuracy was about 80%. OCT-A data did not show significant differences. The feasibility and usefulness of the AI in analyzing FLIO and OCT-A data without any apparent retinal diseases were demonstrated. Although further studies with larger datasets are necessary to validate the results, the results greatly suggest that AI could be useful in analyzing FLIO-data even from healthy subjects without retinal disease and even with small datasets. AI-assisted FLIO is expected to greatly advance early retinal diagnosis.

We present a novel class of Convolutional Neural Networks called Pre-defined Filter Convolutional Neural Networks (PFCNNs), where all $n\times n$ convolution kernels with $n > 1$ are pre-defined and constant during training. It involves a special form of depthwise convolution operation called a Pre-defined Filter Module (PFM). In the channel-wise convolution part, the $1\times n\times n$ kernels are drawn from a fixed pool of only a few (16) different pre-defined kernels. In the $1\times 1$ convolution part linear combinations of the pre-defined filter outputs are learned. Despite this harsh restriction, complex and discriminative features are learned. These findings provide a novel perspective on the way how information is processed within deep CNNs. We discuss various properties of PFCNNs and prove their effectiveness using the popular datasets Caltech101, CIFAR10, CUB-200-2011, FGVC-Aircraft, Flowers102, and Stanford Cars. Our implementation of PFCNNs is provided on Github https://github.com/Criscraft/PredefinedFilterNetworks.

Zielsetzung Eine adäquate Bildqualität von Röntgenaufnahmen ist die Grundlage für eine diagnostische Beurteilung. In der digitalen Röntgendiagnostik können technische Parameter wie der Kontrast nachbearbeitet werden – nicht aber die Positionierung des Patienten während der Aufnahme. Die auf der korrekten Ausrichtung der Anatomie basierende Bildqualität kann bisher nicht automatisch erfasst werden. Das Ziel dieser Arbeit ist die Anwendung eines KI-basierten Algorithmus zur Qualitätsbewertung von Röntgenaufnahmen des Kniegelenks.

Recent findings have shown that highly over-parameterized Neural Networks generalize without pretraining or explicit regularization. It is achieved with zero training error, i.e., complete over-fitting by memorizing the training data. This is surprising, since it is completely against traditional machine learning wisdom. In our empirical study we fortify these findings in the domain of fine-grained image classification. We show that very large Convolutional Neural Networks with millions of weights do learn with only a handful of training samples and without image augmentation, explicit regularization or pretraining. We train the architectures ResNet018, ResNet101 and VGG19 on subsets of the difficult benchmark datasets Caltech101, CUB_200_2011, FGVCAircraft, Flowers102 and StanfordCars with 100 classes and more, perform a comprehensive comparative study and draw implications for the practical application of CNNs. Finally, we show that a randomly initialized VGG19 with 140 million weights learns to distinguish airplanes and motorbikes with up to 95% accuracy using only 20 training samples per class.

Sentiment analysis of news headlines is an important factor that investors consider when making investing decisions. We claim that the sentiment analysis of financial news headlines impacts stock market values. Hence financial news headline data are collected along with the stock market investment data for a period of time. Using Valence Aware Dictionary and Sentiment Reasoning (VADER) for sentiment analysis, the correlation between the stock market values and sentiments in news headlines is established. In our experiments, the data on stock market prices are collected from Yahoo Finance and Kaggle. Financial news headlines are collected from the Wall Street Journal, Washington Post, and Business-Standard website. To cope with such a massive volume of data and extract useful information, various embedding methods, such as Bag-of-words (BoW) and Term Frequency-Inverse Document Frequency (TF-IDF), are employed. These are then fed into machine learning models such as Naive Bayes and XGBoost as well as deep learning models such as Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). Various natural language processing, andmachine and deep learning algorithms are considered in our study to achieve the desired outcomes and to attain superior accuracy than the current state-of-the-art. Our experimental study has shown that CNN (80.86%) and LSTM (84%) are the best performing models in relation to machine learning models, such as Support Vector Machine (SVM) (50.3%), Random Forest (67.93%), and Naive Bayes (59.79%). Moreover, two novel methods, BERT and RoBERTa, were applied with the expectation of better performance than all the other models, and they did exceptionally well by achieving an accuracy of 90% and 88%, respectively.

Within the last decade Deep Learning has become a tool for solving challenging problems like image recognition. Still, Convolutional Neural Networks (CNNs) are considered black-boxes, which are difficult to understand by humans. Hence, there is an urge to visualize CNN architectures, their internal processes and what they actually learn. Previously, virtual realityhas been successfully applied to display small CNNs in immersive 3D environments. In this work, we address the problem how to feasibly render large-scale CNNs, thereby enabling the visualization of popular architectures with ten thousands of feature maps and branches in the computational graph in 3D. Our software "DeepVisionVR" enables the user to freely walk through the layered network, pick up and place images, move/scale layers for better readability, perform feature visualization and export the results. We also provide a novel Pytorch module to dynamically link PyTorch with Unity, which gives developers and researchers a convenient interface to visualize their own architectures. The visualization is directly created from the PyTorch class that defines the Pytorch model used for training and testing. This approach allows full access to the network's internals and direct control over what exactly is visualized. In a use-case study, we apply the module to analyze models with different generalization abilities in order to understand how networks memorize images. We train two recent architectures, CovidResNet and CovidDenseNet on the Caltech101 and the SARS-CoV-2 datasets and find that bad generalization is driven by high-frequency features and the susceptibility to specific pixel arrangements, leading to implications for the practical application of CNNs. The code is available on Github https://github. com/Criscraft/DeepVisionVR.

Structural MRI studies in first-episode psychosis (FEP) and in clinical high risk (CHR) patients have consistently shown volumetric abnormalities in frontal, temporal, and cingulate cortex areas. The aim of the present study was to employ chaos analysis for the identification of brain topology differences in people with psychosis. Structural MRI were acquired from 77 FEP, 73 CHR and 44 healthy controls (HC). Chaos analysis of the gray matter distribution was performed: First, the distances of each voxel from the center of mass in the gray matter image was calculated. Next, the distances multiplied by the voxel intensity were represented as a spatial-series, which then was analyzed by extracting the Largest-Lyapunov-Exponent (lambda). The lambda brain map depicts thus how the gray matter topology changes. Between-group differences were identified by (a) comparing the lambda brain maps, which resulted in statistically significant differences in FEP and CHR compared to HC; and (b) matching the lambda series with the Morlet wavelet, which resulted in statistically significant differences in the scalograms of FEP against CHR and HC. The proposed framework using spatial-series extraction enhances the between-group differences of FEP, CHR and HC subjects, verifies diagnosis-relevant features and may potentially contribute to the identification of structural biomarkers for psychosis.

Feature-product networks (FP-nets) are inspired by end-stopped cortical cells with FP-units that multiply the outputs of two filters. We enhance state-of-the-art deep networks, such as the ResNet and MobileNet, with FP-units and show that the resulting FP-nets perform better on the Cifar-10 and ImageNet benchmarks. Moreover, we analyze the hyperselectivity of the FP-net model neurons and show that this property makes FP-nets less sensitive to adversarial attacks and JPEG artifacts. We then show that the learned model neurons are end-stopped to different degrees and that they provide sparse representations with an entropy that decreases with hyperselectivity.

Cortical slow oscillations (SOs) and thalamocortical sleep spindles are two prominent EEG rhythms of slow wave sleep. These EEG rhythms play an essential role in memory consolidation. In humans, sleep spindles are categorized into slow spindles (8-12 Hz) and fast spindles (12-16 Hz), with different properties. Slow spindles that couple with the up-to-down phase of the SO require more experimental and computational investigation to disclose their origin, functional relevance and most importantly their relation with SOs regarding memory consolidation. To examine slow spindles, we propose a biophysical thalamocortical model with two independent thalamic networks (one for slow and the other for fast spindles). Our modeling results show that fast spindles lead to faster cortical cell firing, and subsequently increase the amplitude of the cortical local field potential (LFP) during the SO down-to-up phase. Slow spindles also facilitate cortical cell firing, but the response is slower, thereby increasing the cortical LFP amplitude later, at the SO up-to-down phase of the SO cycle. Neither the SO rhythm nor the duration of the SO down state is affected by slow spindle activity. Furthermore, at a more hyperpolarized membrane potential level of fast thalamic subnetwork cells, the activity of fast spindles decreases, while the slow spindles activity increases. Together, our model results suggest that slow spindles may facilitate the initiation of the following SO cycle, without however affecting expression of the SO Up and Down states.

Zielsetzung Die korrekte Ausrichtung der Patientenanatomie vor der Röntgenröhre ist elementar für die diagnostische Qualität des resultierenden Röntgenbildes. Tiefenbild (Time-of-Flight

Coronavirus disease (COVID-19) is a viral infection caused by SARS-CoV-2. The modalities such as computed tomography (CT) have been successfully utilized for the early stage diagnosis of COVID-19 infected patients. Recently, many researchers have utilized deep learning models for the automated screening of COVID-19 suspected cases. An ensemble deep learning and Internet of Things (IoT) based framework is proposed for screening of COVID-19 suspected cases. Three well-known pretrained deep learning models are ensembled. The medical IoT devices are utilized to collect the CT scans, and automated diagnoses are performed on IoT servers. The proposed framework is compared with thirteen competitive models over a four-class dataset. Experimental results reveal that the proposed ensembled deep learning model yielded 98.98% accuracy. Moreover, the model outperforms all competitive models in terms of other performance metrics achieving 98.56% precision, 98.58% recall, 98.75% F-score, and 98.57% AUC. Therefore, the proposed framework can improve the acceleration of COVID-19 diagnosis.