Deep Learning In Computer Vision Research Articles

Synthetic Data for Deep Learning in Computer Vision & Medical Imaging: A Means to Reduce Data Bias

Deep-learning (DL) performs well in computer-vision and medical-imaging automated decision-making applications. A bottleneck of DL stems from the large amount of labelled data required to train accurate models that generalise well. Data scarcity and imbalance are common problems in imaging applications that can lead DL models towards biased decision making. A solution to this problem is synthetic data. Synthetic data is an inexpensive substitute to real data for improved accuracy and generalisability of DL models. This survey reviews the recent methods published in relation to the creation and use of synthetic data for computer-vision and medical-imaging DL applications. The focus will be on applications that utilised synthetic data to improve DL models by either incorporating an increased diversity of data that is difficult to obtain in real life, or by reducing a bias caused by class imbalance. Computer-graphics software and generative networks are the most popular data generation techniques encountered in the literature. We highlight their suitability for typical computer-vision and medical-imaging applications, and present promising avenues for research to overcome their computational and theoretical limitations.

Augmenting the Softmax with Additional Confidence Scores for Improved Selective Classification with Out-of-Distribution Data

AbstractDetecting out-of-distribution (OOD) data is a task that is receiving an increasing amount of research attention in the domain of deep learning for computer vision. However, the performance of detection methods is generally evaluated on the task in isolation, rather than also considering potential downstream tasks in tandem. In this work, we examine selective classification in the presence of OOD data (SCOD). That is to say, the motivation for detecting OOD samples is to reject them so their impact on the quality of predictions is reduced. We show under this task specification, that existing post-hoc methods perform quite differently compared to when evaluated only on OOD detection. This is because it is no longer an issue to conflate in-distribution (ID) data with OOD data if the ID data is going to be misclassified. However, the conflation within ID data of correct and incorrect predictions becomes undesirable. We also propose a novel method for SCOD, Softmax Information Retaining Combination (SIRC), that augments a softmax-based confidence score with a secondary class-agnostic feature-based score. Thus, the ability to identify OOD samples is improved without sacrificing separation between correct and incorrect ID predictions. Experiments on a wide variety of ImageNet-scale datasets and convolutional neural network architectures show that SIRC is able to consistently match or outperform the baseline for SCOD, whilst existing OOD detection methods fail to do so. Interestingly, we find that the secondary scores investigated for SIRC do not consistently improve performance on all tested OOD datasets. To address this issue, we further extend SIRC to incorporate multiple secondary scores (SIRC+). This further improves SCOD performance, both generally, and in terms of consistency over diverse distribution shifts. Code is available at https://github.com/Guoxoug/SIRC.

CHD-CXR: a de-identified publicly available dataset of chest x-ray for congenital heart disease.

Congenital heart disease is a prevalent birth defect, accounting for approximately one-third of major birth defects. The challenge lies in early detection, especially in underdeveloped medical regions where a shortage of specialized physicians often leads to oversight. While standardized chest x-rays can assist in diagnosis and treatment, their effectiveness is limited by subtle cardiac manifestations. However, the emergence of deep learning in computer vision has paved the way for detecting subtle changes in chest x-rays, such as lung vessel density, enabling the detection of congenital heart disease in children. This highlights the need for further investigation. The lack of expert-annotated, high-quality medical image datasets hinders the progress of medical image artificial intelligence. In response, we have released a dataset containing 828 DICOM chest x-ray files from children with diagnosed congenital heart disease, alongside corresponding cardiac ultrasound reports. This dataset emphasizes complex structural characteristics, facilitating the transition from machine learning to machine teaching in deep learning. To ascertain the dataset's applicability, we trained a preliminary model and achieved an area under the receiver operating characteristic curve (ROC 0.85). We provide detailed introductions and publicly available datasets at: https://www.kaggle.com/competitions/congenital-heart-disease.

Hyperbolic Deep Learning in Computer Vision: A Survey

AbstractDeep representation learning is a ubiquitous part of modern computer vision. While Euclidean space has been the de facto standard manifold for learning visual representations, hyperbolic space has recently gained rapid traction for learning in computer vision. Specifically, hyperbolic learning has shown a strong potential to embed hierarchical structures, learn from limited samples, quantify uncertainty, add robustness, limit error severity, and more. In this paper, we provide a categorization and in-depth overview of current literature on hyperbolic learning for computer vision. We research both supervised and unsupervised literature and identify three main research themes in each direction. We outline how hyperbolic learning is performed in all themes and discuss the main research problems that benefit from current advances in hyperbolic learning for computer vision. Moreover, we provide a high-level intuition behind hyperbolic geometry and outline open research questions to further advance research in this direction.

A Model for Estimating the Economic Costs of Computer Vision Systems That Use Deep Learning

Deep learning, the most important subfield of machine learning and artificial intelligence (AI) over the last decade, is considered one of the fundamental technologies underpinning the Fourth Industrial Revolution. But despite its record-breaking history, deep learning’s enormous appetite for compute and data means that sometimes it can be too costly to practically use. In this paper, we connect technical insights from deep learning scaling laws and transfer learning with the economics of IT to propose a framework for estimating the cost of deep learning computer vision systems to achieve a desired level of accuracy. Our tool can be of practical use to AI practitioners in industry or academia to guide investment decisions.

Deep Learning-Based Small Face Detection from Hard Image

Facial detection usually comes first in face recognition and face analysis systems. Previously, techniques such as directed gradient histograms and cascades relied on manually-engineered features from particular photos. Nevertheless, the precision with which these techniques could identify faces in uncontrolled environments was restricted. Numerous deep learning-based face recognition frameworks have recently been developed, many of which have significantly increased accuracy, as a result of the rapid progress of deep learning in computer vision. Despite these advancements, detecting small, scaled, positioned, occluded, blurred, and faces that are partially occluded in uncontrolled conditions remains a challenge in face identification. This problem has been studied for many years but has not been completely resolved.

A Short Review on Supervised Machine Learning and Deep Learning Techniques in Computer Vision

In last years, computer vision has shown important advances, mainly using the application of supervised machine learning (ML) and deep learning (DL) techniques. The objective of this review is to show a brief review of the current state of the field of supervised ML and DL techniques, especially on computer vision tasks. This study focuses on the main ideas, advantages, and applications of DL in computer vision and highlights their main concepts and advantages. This study showed the strengths, limitations, and effects of computer vision supervised ML and DL techniques.

Evaluation of deep learning computer vision for water level measurements in rivers

Image-based gauging stations offer the potential for substantial enhancement in the monitoring networks of river water levels. Nonetheless, the majority of camera gauges fall short in delivering reliable and precise measurements because of the fluctuating appearance of water in the rivers over the course of the year. In this study, we introduce a method for measuring water levels in rivers using both the traditional continuous image subtraction (CIS) approach and a SegNet neural network based on deep learning computer vision. The historical images collected from on-site investigations were employed to train three neural networks (SegNet, U-Net, and FCN) in order to evaluate their effectiveness, overall performance, and reliability. The research findings demonstrated that the SegNet neural network outperformed the CIS method in accurately measuring water levels. The root mean square error (RMSE) between the water level measurements obtained by the SegNet neural network and the gauge station's readings ranged from 0.013 m to 0.066 m, with a high correlation coefficient of 0.998. Furthermore, the study revealed that the performance of the SegNet neural network in analyzing water levels in rivers improved with the inclusion of a larger number of images, diverse image categories, and higher image resolutions in the training dataset. These promising results emphasize the potential of deep learning computer vision technology, particularly the SegNet neural network, to enhance water level measurement in rivers. Notably, the quality and diversity of the training dataset play a crucial role in optimizing the network's performance. Overall, the application of this advanced technology holds great promise for advancing water level monitoring and management in river systems.

Using deep learning to enhance electronic service quality: Application to real estate websites

Electronic service quality (E-SQ) is a strategic metric for successful e-services. Among the service quality dimensions, tangibility is overlooked. However, by incorporating visuals or tangible tools, the intangible nature of e-services can be balanced. Thanks to advancements in Deep Learning for computer vision, tangible visual features can now be leveraged to enhance the browsing and searching experience electronic services. Users usually have specific search criteria to meet, but most services won't offer flexible search filters. This research emphasizes the importance of integrating visual and descriptive features to improve the tangibility and efficiency of e-services. A prime example of an electronic service that can benefit from this is real-estate websites. Searching for real estate properties that match user preferences is usually demanding and lacks visual filters, such as the Damage Level to the property. The research introduces a novel visual descriptive feature, the Damage Level, which utilizes a deep learning network known as Mask-RCNN to estimate damage in real estate images. Additionally, a model is developed to incorporate the Damage Level as a tangible feature in electronic real estate services, with the aim of enhancing the tangible customer experience.

Detection of Bowel and Mesenteric Injuries Using Deep Learning Computer Vision Models

Background/Objective:While only seen in 1-5% of patients who undergo a CT (computed tomography) scan, blunt bowel and mesenteric injuries (BMI) are associated with significantly increased morbidity and mortality. A significant cause of the increased morbidity of BMI is due to the difficulty of diagnosis from clinical and imaging information which leads to delay in diagnosis. Accurate and timely diagnosis is vital to reduce the morbidity of BMI.
 Methods:For this project, our primary objective is to create a binary prediction model that determines if a patient has BMI based on their abdominal CT scans. Due to the importance of the early and definitive diagnosis of BMI in trauma patients, an extension of this project will seek to introduce explainability into the model to highlight which features on the CT scan caused the model to make its prediction. The patients with BMI were sourced from a trauma registry that recorded trauma cases from IU Health with relevant diagnosis codes. The images from our search will be reduced to the relevant slices for diagnosis of BMI and then used to train an ML model to makea yes/no prediction from the image. Once the model is trained, testing data will be evaluated on the model and the gradient vectors from the model during inference will be used to create a heatmap with GRAD-CAM that illustrates what portions of the image were relevant for the decision made by the algorithm.
 Future Directions:Using the collected abdominal CTs, we can train our machine learning pipeline to detect BMI. Based on the performance of the model, we will determine if we need to collect more data. Then, we can evaluate the explainability of the model using GRAD-CAM and compareperformance of the ML model to the performance of expert and trainee radiologists.

An AI-assisted integrated, scalable, single-cell phenomic-transcriptomic platform to elucidate intratumor heterogeneity against immune response.

We present a novel framework combining single-cell phenotypic data with single-cell transcriptomic analysis to identify factors underpinning heterogeneity in antitumor immune response. We developed a pairwise, tumor-immune discretized interaction assay between natural killer (NK-92MI) cells and patient-derived head and neck squamous cell carcinoma (HNSCC) cell lines on a microfluidic cell-trapping platform. Furthermore we generated a deep-learning computer vision algorithm that is capable of automating the acquisition and analysis of a large, live-cell imaging data set (>1 million) of paired tumor-immune interactions spanning a time course of 24 h across multiple HNSCC lines (n = 10). Finally, we combined the response data measured by Kaplan-Meier survival analysis against NK-mediated killing with downstream single-cell transcriptomic analysis to interrogate molecular signatures associated with NK-effector response. As proof-of-concept for the proposed framework, we efficiently identified MHC class I-driven cytotoxic resistance as a key mechanism for immune evasion in nonresponders, while enhanced expression of cell adhesion molecules was found to be correlated with sensitivity against NK-mediated cytotoxicity. We conclude that this integrated, data-driven phenotypic approach holds tremendous promise in advancing the rapid identification of new mechanisms and therapeutic targets related to immune evasion and response.

A survey of multimodal hybrid deep learning for computer vision: Architectures, applications, trends, and challenges

In recent years, deep learning algorithms have rapidly revolutionized artificial intelligence, particularly machine learning, enabling researchers and practitioners to extend previously hand-crafted feature extraction procedures. In particular, deep learning uses adaptive learning processes to learn more complex and informative patterns from datasets of varying sizes. With the increasing availability of multimodal data streams and recent advances in deep learning algorithms, multimodal deep learning is on the rise. This requires the development of complex models that can process and analyze multimodal information in a consistent manner. However, unstructured data can come in many different forms (also known as modalities). Extracting relevant features from this data remains an ambitious goal for deep learning researchers. According to the literature, most deep learning systems consist of a single architecture (i.e., standalone deep learning). When two or more deep learning architectures are combined over multiple sensory modalities, the result is called a multimodal hybrid deep learning model. Since this research direction has received much attention in the field of deep learning, the purpose of this survey is to provide a broader overview of the topic. In this paper, we provide a comprehensive review of recent advances in multimodal hybrid deep learning, including a thorough analysis of the most commonly developed hybrid architectures. In particular, one of the main challenges in multimodal hybrid analysis is the ability of these architectures to systematically integrate cross-modal features in hybrid designs. Therefore, we propose a generic framework for multimodal hybrid learning that focuses mainly on fusion methods. We also identify trends and challenges in multimodal hybrid learning and provide insights and directions for future research. Our findings show that multimodal hybrid learning can perform well in a variety of challenging computer vision applications and tasks.

Harnessing Deep Learning in Computer Vision for Effective Face Mask Detection in the Global COVID-19 Crisis

Objective: This study aims to explore the use of deep learning algorithms in computer vision to address the effective detection of face mask use in the global COVID-19 crisis.
 Results: Deep learning techniques enable the execution of tasks by computers and smart devices without human intervention, including image identification and predictions. Its applications have demonstrated significant advances in several areas, particularly in engineering and health research. The study highlights that computer vision scientists can contribute to the prevention, control and management of the fight against COVID-19 and other airborne viruses. Computer vision employs algorithmic tools to process images, perform associations and transmit relevant information.
 Implications for Research, Practice and Social: Construction of a deep learning model using a dataset of people with and without face masks. The developed model, implemented in Python with the help of OpenCV, Keras and TensorFlow libraries, presented highly promising results, reaching an accuracy of 99% in predictions in test images.
 Originality/Value: This study highlights the originality and value of deep learning techniques in computer vision as an effective means of tackling virus-borne pandemics such as COVID-19, and contributing to a preventive, efficient and cost-effective approach to use of face masks.
 Keywords: Computer vision; Convolutional Neural Network; Face mask.

Leveraging an AI-Assisted Electrochemical Sensor Array to Visualize Digital Fingerprints of Complex Liquids

Electrochemical sensor arrays have shown potential for fast and untargeted chemical analysis of multi-component media, enabling simultaneous quantification of multiple analytes and estimation of liquid properties that could correlate with human sensory perception. In particular, potentiometric electronic tongues (e-tongues) have been explored as promising alternative tools for chemical analysis in various applications, ranging from traceability of goods to characterization of food products for quality control and innovation processes. The reduced complexity of non-selective sensor fabrication and the ease of potentiometric transduction make these devices suitable for portable and decentralized chemical analysis. However, the interpretation of the sensor array response has always been a challenge due to the inherent cross-sensitivity of the sensing materials and the corresponding combinatorial signals arising upon sensor interaction with a liquid. Machine learning can help recognizing and mapping signal patterns and various data processing, unsupervised and supervised techniques have been proposed to calibrate e-tongue devices. These methods usually require exposing the sensor array to a set of liquids with known properties that serve as training base to build a calibration model. Therefore, the quality and quantity of tests is crucial to boost sensor performances. Nevertheless, performing an extensive number of measurements could be extremely time-consuming and results to be tedious to achieve in practice for certain use-cases. In this context, advances in deep learning and machine learning models have shown potential to accelerate chemical and materials discovery when combined with high-throughput experimentation, highlighting the benefits of AI-assisted research practices. Moreover, the recent advent of multi-domain and multi-task models trained by self-supervision, so-called foundation models, bears also promises for extending learnt representations across multiple fields, thus counteracting the reduced data availability in certain applications, and benefiting from information exchange across domains. Thus, in the present contribution we propose extending this approach to data-driven chemical sensors. More specifically, we leverage transfer learning based on fingerprints pretrained in other domains to model new instrument/sensor data representations. Herein, we demonstrate how the output of a model system comprising an integrated electrochemical sensor array for analysis of multi-component liquids can be encoded as image representations to leverage existing deep learning computer vision models pretrained on large collections of image data. The models effectively extract features from these representations and feed specific model heads to perform downstream tasks. Firstly, an integrated sensor array comprising 16 polymeric sensors was fabricated through electrodeposition on conventional electroless nickel immersion gold (ENIG) electrodes. The conductive polymers (PEDOT, PPy, PANI and PAPBA) were synthesized by chronoamperometry or cyclic voltammetry in a three-electrode configuration and were enriched with doping agents for enhanced sensitivity. 15 linearly independent differential voltages between these polymeric sensors were measured during the transition of the sensor array from a reference solution to a test solution, thereby obviating the need for a conventional reference electrode. Indeed, the use of low-selective polymeric sensors for potentiometric measurements does not necessarily require integration of reference electrodes, which are known to be unpractical for remote sensing applications. Training data were obtained by alternatively immersing the sensor array in reference (120 s) and test (60 s) solutions continuously using an automated test rig. The 15 raw time-series data from the sensor array were processed and concatenated to yield a spectral response, which was smoothed by means of the moving average technique and standardized using Standard Normal Variate (SNV). The obtained spectra were encoded into image representations using the Gramian Angular Summation Field transformation. Off-the-shelf features were generated leveraging pretrained neural networks developed to classify natural images and applied to the “sensor images”. Dimensionality reduction through Principal Component Analysis (PCA) yielded a set of features that could then be used to train machine learning classifiers and regressors. The pipeline was applied to generate visual fingerprints of multiple beverages, proving full discrimination of liquid types (mineral waters, coffees, fruit juices, soft beverages and wines). On a model dataset comprising 11 Italian red wines, it was demonstrated that image fingerprints of samples enabled class identification with a mean accuracy ~95%. The results demonstrate the successful creation of a new representation of the chemical sensing space which achieves comparable performance as domain-specific hand-crafted feature selection. The present contribution represents an example of integration of data processing techniques and publicly available libraries/models to support transfer of methodologies across domains. We believe this approach could be disruptive in the field of electrochemical sensor arrays, especially for processing e-tongues and e-noses response and enhancing capabilities of data-driven chemical sensors. Figure 1

Physical distancing and its association with travel behavior in daily pre-pandemic urban life: An analysis utilizing lifelogging images and composite survey and mobility data

This study analyzed physical distancing in people’s daily lives and its association with travel behavior and the use of transportation modes before the COVID-19 outbreak. We used data from photographic images acquired automatically by lifelogging devices every 5 seconds, on average, from 170 participants of a 2-day wearable camera study, in order to identify their physical distancing status throughout the day. Using deep-learning computer vision algorithms, we developed three measures which provided a near-continuous quantification of the proportion of time spent without anyone else within a distance of approximately 13 meters, as well as the proportion of time spent without others within approximately 2 meters. These measures are then used as outcomes in beta regression and multinomial logit models to explore the association between the participant’s physical distancing and travel behavior and transportation choices. The multidisciplinary research approach to understand these associations accounted for a number of social, economic, and cultural factors that potentially influenced their physical isolation levels. We found that participants spend a significant amount of time physically separated from others, without anyone else within 2 meters. The use of public transportation, automobiles, active travel, and an increase in trip frequency, including trips to transportation facilities, reduced the extent of physical distancing, with public transportation having the most significant impact. Higher incomes, strong social networks, and a sense of belonging to the community reduced the tendency for physical distancing. In contrast, factors such as age, obesity, dog ownership, intensive use of the Internet, and being knowledgeable about climate change issues increased the likelihood of physical distancing. The paper addresses a crucial gap in our understanding of how these factors intersect to create the dynamics of physical distancing in non-emergency situations and highlights their planning and operational implications while showcasing the use of unique person-based physical distancing measures derived from autonomously collected image data.

The application of deep learning in computer vision: Facial emotion recognition based on convolutional neural network

Emotion recognition is an essential task in many fields, including affective computing, psychology, and human-computer interaction. The ability to accurately recognize emotions from facial expressions can help in the development of more personalized and responsive systems. The objective of this paper is to investigate the effectiveness of a Convolutional Neural Network (CNN) model in classifying emotions from facial expressions. We used the images classified into seven different categories from FER2013 dataset. The model is implemented using TensorFlow Keras, which consists of four CNN layers followed by two fully connected (FCN) layers and an output layer. In between each layer we use batch normalization, max pooling, and dropout to reduce calculation cost and prevent overfitting. Our model achieved an accuracy of 66.85% and a precision accuracy of 77.18% on the training set after 20 epochs. The recall and the F-1 score is 56.93% and 0.6545 respectively. This study demonstrates that the proposed CNN model can effectively recognize emotions in facial images. The availability of the FER2013 dataset provides researchers with an opportunity to further explore this area of research. The proposed model can be useful in various applications, such as affective computing, human-computer interaction, and psychology. Future work could involve tuning the hyperparameters of the model or using different datasets to explore the generalizability of the proposed model. Overall, our study provides insights into the effectiveness of CNNs for emotion recognition and highlights the potential for further research in this area.

Deep Learning Artificial Intelligence Tool for Automated Radiographic Determination of Posterior Tibial Slope in Patients With ACL Injury.

An increased posterior tibial slope (PTS) corresponds with an increased risk of graft failure after anterior cruciate ligament (ACL) reconstruction (ACLR). Validated methods of manual PTS measurements are subject to potential interobserver variability and can be inefficient on large datasets. To develop a deep learning artificial intelligence technique for automated PTS measurement from standard lateral knee radiographs. It was hypothesized that this deep learning tool would be able to measure the PTS on a high volume of radiographs expeditiously and that these measurements would be similar to previously validated manual measurements. Cohort study (diagnosis); Level of evidence, 2. A deep learning U-Net model was developed on a cohort of 300 postoperative short-leg lateral radiographs from patients who underwent ACLR to segment the tibial shaft, tibial joint surface, and tibial tuberosity. The model was trained via a random split after an 80 to 20 train-validation scheme. Masks for training images were manually segmented, and the model was trained for 400 epochs. An image processing pipeline was then deployed to annotate and measure the PTS using the predicted segmentation masks. Finally, the performance of this combined pipeline was compared with human measurements performed by 2 study personnel using a previously validated manual technique for measuring the PTS on short-leg lateral radiographs on an independent test set consisting of both pre- and postoperative images. The U-Net semantic segmentation model achieved a mean Dice similarity coefficient of 0.885 on the validation cohort. The mean difference between the human-made and computer-vision measurements was 1.92° (σ = 2.81° [P = .24]). Extreme disagreements between the human and machine measurements, as defined by ≥5° differences, occurred <5% of the time. The model was incorporated into a web-based digital application front-end for demonstration purposes, which can measure a single uploaded image in Portable Network Graphics format in a mean time of 5 seconds. We developed an efficient and reliable deep learning computer vision algorithm to automate the PTS measurement on short-leg lateral knee radiographs. This tool, which demonstrated good agreement with human annotations, represents an effective clinical adjunct for measuring the PTS as part of the preoperative assessment of patients with ACL injuries.

Emotion Recognition with CNN

Abstract: Emotion is a subjective phenomenon, utilizing knowledge and science behind tagged data and extracting the components that comprise it has been a difficult challenge. With the advancement of deep learning in computer vision, emotion identification has become a popular research topic. This Project presents feature extraction of facial expressions using a neural network combination for the recognition of various facial emotions (sad, happy, neutral, angry, surprised, fear). Convolution Neural Network has been used to achieve a accuracy of 75%, which have excellent recognition of image features. Haar-Cascade has been used to find the region that contains the face, so the model has to only work with the region which has face.

Artificial Intelligence Derived Changes between Baseline and Interim FDG-PET/CT Radiomics Features Are Associated with Survival Outcomes in Diffuse Large B-Cell Lymphoma (DLBCL)

Background: Relapsed DLBCL occurs in up to 40% of patients and has a poor prognosis. Current prognostic tools such as the international prognostic index (IPI) and newer forms (NCCN-IPI) only include baseline clinical features. PET/CT features have not yet been incorporated into prognostic risk models despite being essential for staging and response assessment. Deep learning computer algorithms have consistently demonstrated the ability to analyze imaging data and identify features that are not readily apparent to the human eye. Using a deep learning computer vision model that automatically extracted radiomic features from PET/CTs, we examined the association of changes between baseline PET and interim PET (PET2) radiomic features on survival outcomes and compared the prognostic performance with NCCN-IPI. Methods: Adult patients with newly diagnosed diffuse large B-cell lymphoma (DLBCL) between 2005 and 2015 and enrolled in the prospective MER cohort were included. Native DICOM images from FDG PET/CT scans at baseline and interim PET (PET2, after 2-3 cycles of chemotherapy) were preprocessed for regularity and then segmented using the nnU-Net architecture with an external lymphoma-specific model. Feature extraction and standard uptake value (SUV) characterization were accomplished with PyRadiomics and NiBabel. Our computational method included no manual segmentation correction. Individual radiomics features (n=115) and the % change above (Δ High) and below (Δ low) 50 th quartile from baseline to PET2 were evaluated for association with event free survival (EFS) and overall survival (OS) using a univariate Cox regression model and by Kaplan Meier analysis. EFS was defined as time from PET2 to first systemic relapse or death from any cause. OS was defined as time from PET2 to death from any cause. Additionally, Spearman and Pearson correlation matrices were used to guide feature selection while minimizing collinearity. We then assessed the association of the highest scoring radiomics features with EFS in a multivariate cox regression model. Results: A total of 506 patients had baseline and PET2 scans available for radiomic feature extraction. Median age at diagnosis was 61 years (range 18-90), 43% were females. NCCN-IPI composition and outcomes of our cohort was comparable to the previously published NCCN cohort, with patient distribution and 5-year OS as follows: Low (9%, 5-y OS 83%), Low-intermediate (38%, 5-y OS 49%), High-intermediate (41%, 5-y OS 31%), High (11%, 5-y OS 12%). At a median follow up of 81 months,there were 308 events, and272 patients had died.Several radiomics features were associated with EFS including change in tumor surface area (SA Δ) (HR: 1.75, 95% CI 1.36-2.24, p&amp;lt;0.005, median of 26,927 mm 2, Figure 1A) and change in total metabolic tumor volume (TMTV Δ) (HR: 1.64, 95% CI 1.27-2.10, p&amp;lt;0.005, median of 120,473 mm 3, Figure 1B). As a reference, NCCN-IPI univariate Cox regression analysis of EFS produced a Harrell's C-statistic (c) of 0.624. Single radiomics features of SA Δ and TMTV Δ individually produced unadjusted c-statistics of 0.616 and 0.614 respectively. Combinations of radiomics features (SA Δ, TMTV Δ and NCCN-IPI classification) demonstrated cumulative benefit (c=0.652). Conclusion: Automatic PET-CT radiomic feature extraction using deep learning models is feasible among patients with newly diagnosed DLBCL. The change in tumor SA and TMTV appear to be promising radiomic biomarkers. A low decline in either tumor SA or TMTV by PET2 is associated with higher risk of relapse and poor survival outcomes. As individual characteristics, these features are similar prognostic indicators of EFS as the NCCN-IPI. Further investigation into risk stratification when combining radiomics with clinical and genomic variables is warranted. Figure 1. Event free survival for DLBCL patients with A) a change in tumor surface area by PET2 less than 50% percentile (SA_Δ Low) and greater than 50 th percentile (SA_Δ High) and B) a change in TMTV by PET2 less than 50% percentile (TMTV_Δ Low) and greater than 50 th percentile (TMTV_Δ High).

Automated quality evaluation for laser cutting in lithium metal battery production using an instance segmentation convolutional neural network

Separating lithium metal foil into individual anodes is a critical process step in all-solid-state battery production. With the use of nanosecond-pulsed laser cutting, a characteristic quality-decisive cut edge geometry is formed depending on the chosen parameter set. This cut edge can be characterized by micrometer-scale imaging techniques such as confocal laser scanning microscopy. Currently, experimental determination of suitable process parameters is time-consuming and biased by the human measurement approach, while no methods for automated quality assurance are known. This study presents a deep-learning computer vision approach for geometry characterization of lithium foil laser cut edges. The convolutional neural network architecture Mask R-CNN was implemented and applied for categorizing confocal laser scanning microscopy images showing defective and successful cuts, achieving a classification precision of more than 95%. The algorithm was trained for automatic pixel-wise segmentation of the quality-relevant melt superelevation along the cut edge, reaching segmentation accuracies of up to 88%. Influence of the training data set size on the classification and segmentation accuracies was assessed confirming the algorithm’s industrial application potential due to the low number of 246 or fewer original images required. The segmentation masks were combined with topography data of cut edges to obtain quantitative metrics for the quality evaluation of lithium metal electrodes. The presented computer vision pipeline enables the integration of an automated image evaluation for quality inspection of lithium foil laser cutting, promoting industrial production of all-solid-state batteries with lithium metal anode.