Diverse Training Datasets Research Articles

Advanced Algorithms for Offline Signature Recognition and Enhanced Forgery Detection Mechanisms

In contemporary society, signatures hold significant importance in various critical documents such as bank cheques, passports, and driver's licenses. Unfortunately, they can be counterfeited through various means, giving rise to issues like fraudulent identifications, identity theft, and cyber-attacks. To mitigate this problem, our project is centered around the development of a system that discerns the authenticity of a signature, distinguishing between genuine and forged signatures within a dataset. We have employed Convolutional Neural Networks (CNN) and deep learning for this purpose. This choice is driven by the understanding that signatures evolve over time due to a range of behavioral factors like age, mental state, and physical well-being. Consequently, our system is designed to adapt and learn from diverse training datasets, enhancing its accuracy in detection. While online and offline signature verification methods exist, our project primarily focuses on the latter, specifically targeting the detectionof forged offline signatures.

Ethical Implications of AI-Driven AAC Systems: Ensuring Inclusivity and Equity in Assistive Technologies

This study investigates the ethical implications of Artificial Intelligence (AI) integration into Augmentative and Alternative Communication (AAC) systems, focusing on inclusivity, accessibility, and data privacy in assistive technologies. AI-driven AAC systems enhance communication for individuals with impairments by offering features like predictive text, speech recognition, and symbol-to-text translation. However, ethical challenges arise as these systems increasingly rely on sensitive user data, which could jeopardize privacy without strict safeguards. Furthermore, limitations in linguistic diversity within AI training data compromise inclusivity, potentially excluding users from varied cultural and linguistic backgrounds. Accessibility also remains an issue, as AI-driven AAC systems are often financially inaccessible for marginalized communities, exacerbating social inequities. Through a comprehensive analysis of current systems, including Proloquo2Go, Tobii Dynavox and Google’s Project Euphonia, this paper evaluates both the advancements and ethical shortcomings in AI-driven AAC technologies. The findings advocate for privacy-preserving AI practices, culturally diverse training datasets, and affordable AAC solutions to foster a more inclusive, accessible, and equitable digital communication environment.

Comparative Analysis of the Response Accuracies of Large Language Models in the Korean National Dental Hygienist Examination Across Korean and English Questions.

Large language models such as Gemini, GPT-3.5, and GPT-4 have demonstrated significant potential in the medical field. Their performance in medical licensing examinations globally has highlighted their capabilities in understanding and processing specialized medical knowledge. This study aimed to evaluate and compare the performance of Gemini, GPT-3.5, and GPT-4 in the Korean National Dental Hygienist Examination. The accuracy of answering the examination questions in both Korean and English was assessed. This study used a dataset comprising questions from the Korean National Dental Hygienist Examination over 5 years (2019-2023). A two-way analysis of variance (ANOVA) test was employed to investigate the impacts of model type and language on the accuracy of the responses. Questions were input into each model under standardized conditions, and responses were classified as correct or incorrect based on predefined criteria. GPT-4 consistently outperformed the other models, achieving the highest accuracy rates across both language versions annually. In particular, it showed superior performance in English, suggesting advancements in its training algorithms for language processing. However, all models demonstrated variable accuracies in subjects with localized characteristics, such as health and medical law. These findings indicate that GPT-4 holds significant promise for application in medical education and standardized testing, especially in English. However, the variability in performance across different subjects and languages underscores the need for ongoing improvements and the inclusion of more diverse and localized training datasets to enhance the models' effectiveness in multilingual and multicultural contexts.

Beyond the stereotypes: Artificial Intelligence image generation and diversity in anesthesiology.

Artificial Intelligence (AI) is increasingly being integrated into anesthesiology to enhance patient safety, improve efficiency, and streamline various aspects of practice. This study aims to evaluate whether AI-generated images accurately depict the demographic racial and ethnic diversity observed in the Anesthesia workforce and to identify inherent social biases in these images. This cross-sectional analysis was conducted from January to February 2024. Demographic data were collected from the American Society of Anesthesiologists (ASA) and the European Society of Anesthesiology and Intensive Care (ESAIC). Two AI text-to-image models, ChatGPT DALL-E 2 and Midjourney, generated images of anesthesiologists across various subspecialties. Three independent reviewers assessed and categorized each image based on sex, race/ethnicity, age, and emotional traits. A total of 1,200 images were analyzed. We found significant discrepancies between AI-generated images and actual demographic data. The models predominantly portrayed anesthesiologists as White, with ChatGPT DALL-E2 at 64.2% and Midjourney at 83.0%. Moreover, male gender was highly associated with White ethnicity by ChatGPT DALL-E2 (79.1%) and with non-White ethnicity by Midjourney (87%). Age distribution also varied significantly, with younger anesthesiologists underrepresented. The analysis also revealed predominant traits such as "masculine, ""attractive, "and "trustworthy" across various subspecialties. AI models exhibited notable biases in gender, race/ethnicity, and age representation, failing to reflect the actual diversity within the anesthesiologist workforce. These biases highlight the need for more diverse training datasets and strategies to mitigate bias in AI-generated images to ensure accurate and inclusive representations in the medical field.

Pitfalls in Interpretive Applications of Artificial Intelligence in Radiology.

Interpretive artificial intelligence (AI) tools are poised to change the future of radiology. However, certain pitfalls may pose particular challenges for optimal AI interpretative performance. These include anatomic variants, age-related changes, postoperative changes, medical devices, image artifacts, lack of integration of prior and concurrent imaging examinations and clinical information, and the satisfaction-of-search effect. Model training and development should account for such pitfalls to minimize errors and optimize interpretation accuracy. More broadly, AI algorithms should be exposed to diverse and complex training datasets to yield a holistic interpretation that considers all relevant information beyond the individual examination. Successful clinical deployment of AI tools will require that radiologist end users recognize these pitfalls and other limitations of the available models. Furthermore, developers should incorporate explainable AI techniques (e.g., heat maps) into their tools, to improve radiologists' understanding of model outputs and to enable radiologists to provide feedback for guiding continuous learning and iterative refinement. In this article, we provide an overview of common pitfalls that radiologists may encounter when using interpretive AI products in daily practice. We present how such pitfalls lead to AI errors and offer potential strategies that AI developers may use for their mitigation.

Classification of Practical Floor Moisture Damage Using GPR - Limits and Opportunities

Machine learning in non-destructive testing (NDT) offers significant potential for efficient daily data analysis and uncovering previously unknown relationships in persistent problems. However, its successful application heavily depends on the availability of a diverse and well-labeled training dataset, which is often lacking, raising questions about the transferability of trained algorithms to new datasets. To examine this issue closely, the authors applied classifiers trained with laboratory Ground Penetrating Radar (GPR) data to categorize on-site moisture damage in layered building floors. The investigations were conducted at five different locations in Germany. For reference, cores were taken at each measurement point and labeled as (i) dry, (ii) with insulation damage, or (iii) with screed damage. Compared to the accuracies of 84 % to 90 % within the laboratory training data (504 B-Scans), the classifiers achieved a lower overall accuracy of 53 % for on-site data (72 B-Scans). This discrepancy is mainly attributable to a significantly higher dynamic of all signal features extracted from on-site measurements compared to laboratory training data. Nevertheless, this study highlights the promising sensitivity of GPR for identifying individual damage cases. In particular the results showing insulation damage, which cannot be detected by any other non-destructive method, revealed characteristic patterns. The accurate interpretation of such results still depends on trained personnel, whereby fully automated approaches would require a larger and diverse on-site data set. Until then, the findings of this work contribute to a more reliable analysis of moisture damage in building floors using GPR and offer practical insights into applying machine learning to non-destructive testing for civil engineering (NDT-CE).

Defog YOLO for road object detection in foggy weather

Abstract Object detection research predominantly focuses on clear weather conditions, often overlooking the challenges posed by foggy weather. Fog impairs the vision of onboard cameras, creating significant obstacles for autonomous vehicles. To tackle these issues, we present the Defog YOLO algorithm, specifically designed for road object detection in foggy conditions. Our approach integrates an enhanced U-Net framework for visual defogging, where the encoder leverages super-resolution back projection to combine multi-layer features. The decoder employs a back projection feedback mechanism to improve image restoration. Additionally, we augment the Feature Pyramid Network with a noise-aware attention mechanism, allowing the network to emphasize critical channel and spatial information while mitigating noise. Given the scarcity of labeled foggy images, we introduce a fog addition module to generate a more diverse training dataset. We validate our method using a synthesized FOG-TRAINVAL dataset, derived from the VOC dataset, demonstrating its robustness in foggy scenarios. Experimental results show that our proposed method achieves an mAP score of 60% on the Real-world Task-driven Testing Set foggy weather test set, with a precision of 86.7% and a recall of 54.2%. These findings underscore the effectiveness and improved generalizability of our approach for object detection in adverse weather conditions.

Controllable Seismic Velocity Synthesis Using Generative Diffusion Models

AbstractAccurate seismic velocity estimations are vital to understanding Earth's subsurface structures, assessing natural resources, and evaluating seismic hazards. Machine learning‐based inversion algorithms have shown promising performance in regional (i.e., for exploration) and global velocity estimations, while their effectiveness hinges on access to large and diverse training data sets whose distributions generally cover the target solutions. Additionally, enhancing the precision and reliability of velocity estimation also requires incorporating prior information, for example, geological classes, well logs, and subsurface structures, but current statistical or neural network‐based methods are not flexible enough to handle such multimodal information. To address both challenges, we propose to use conditional generative diffusion models for seismic velocity synthesis in which we readily incorporate those priors. This approach enables the generation of seismic velocities that closely match the expected target distribution, offering data sets informed by both expert knowledge and measured data to support training for data‐driven geophysical methods. We demonstrate the flexibility and effectiveness of our method through training diffusion models on the OpenFWI data set under various conditions, including class labels, well logs, reflectivity images, and the combination of these priors. The performance of the approach under out‐of‐distribution conditions further underscores its generalization ability, showcasing its potential to provide tailored priors for velocity inverse problems and create specific training data sets for machine learning‐based geophysical applications.

A Survey on Offline Reinforcement Learning: Taxonomy, Review, and Open Problems.

With the widespread adoption of deep learning, reinforcement learning (RL) has experienced a dramatic increase in popularity, scaling to previously intractable problems, such as playing complex games from pixel observations, sustaining conversations with humans, and controlling robotic agents. However, there is still a wide range of domains inaccessible to RL due to the high cost and danger of interacting with the environment. Offline RL is a paradigm that learns exclusively from static datasets of previously collected interactions, making it feasible to extract policies from large and diverse training datasets. Effective offline RL algorithms have a much wider range of applications than online RL, being particularly appealing for real-world applications, such as education, healthcare, and robotics. In this work, we contribute with a unifying taxonomy to classify offline RL methods. Furthermore, we provide a comprehensive review of the latest algorithmic breakthroughs in the field using a unified notation as well as a review of existing benchmarks' properties and shortcomings. Additionally, we provide a figure that summarizes the performance of each method and class of methods on different dataset properties, equipping researchers with the tools to decide which type of algorithm is best suited for the problem at hand and identify which classes of algorithms look the most promising. Finally, we provide our perspective on open problems and propose future research directions for this rapidly growing field.

Improving the radiographic image analysis of the classic metaphyseal lesion via conditional diffusion models

The classic metaphyseal lesion (CML) is a unique fracture highly specific for infant abuse. This fracture is often subtle in radiographic appearance and commonly occurs in the distal tibia. The development of an automated model that can accurately identify distal tibial radiographs with CMLs is important to assist radiologists in detecting these fractures. However, building such a model typically requires a large and diverse training dataset. To address this problem, we propose a novel diffusion model for data augmentation called masked conditional diffusion model (MaC-DM). In contrast to previous generative models, our approach produces a wide range of realistic-appearing synthetic images of distal tibial radiographs along with their associated segmentation masks. MaC-DM achieves this by incorporating weighted segmentation masks of the distal tibias and CML fracture sites as image conditions for guidance. The augmented images produced by MaC-DM significantly enhance the performance of various commonly used classification models, accurately distinguishing normal distal tibial radiographs from those with CMLs. Additionally, it substantially improves the performance of different segmentation models, accurately labeling areas of the CMLs on distal tibial radiographs. Furthermore, MaC-DM can control the size of the CML fracture in the augmented images.

Marker Data Enhancement For Markerless Motion Capture.

Human pose estimation models can measure movement from videos at a large scale and low cost; however, open-source pose estimation models typically detect only sparse keypoints, which leads to inaccurate joint kinematics. OpenCap, a freely available service for researchers to measure movement from videos, addresses this issue using a deep learning model-the marker enhancer-that transforms sparse keypoints into dense anatomical markers. However, OpenCap performs poorly on movements not included in the training data. Here, we create a much larger and more diverse training dataset and develop a more accurate and generalizable marker enhancer. We compiled marker-based motion capture data from 1176 subjects and synthesized 1433 hours of keypoints and anatomical markers to train the marker enhancer. We evaluated its accuracy in computing kinematics using both benchmark movement videos and synthetic data representing unseen, diverse movements. The marker enhancer improved kinematic accuracy on benchmark movements (mean error: 4.1°, max: 8.7°) compared to using video keypoints (mean: 9.6°, max: 43.1°) and OpenCap's original enhancer (mean: 5.3°, max: 11.5°). It also better generalized to unseen, diverse movements (mean: 4.1°, max: 6.7°) than OpenCap's original enhancer (mean: 40.4°, max: 252.0°). Our marker enhancer demonstrates both accuracy and generalizability across diverse movements. We integrated the marker enhancer into OpenCap, thereby offering its thousands of users more accurate measurements across a broader range of movements.

A deep learning-based toolkit for 3D nuclei segmentation and quantitative analysis in cellular and tissue context.

We present a new set of computational tools that enable accurate and widely applicable 3D segmentation of nuclei in various 3D digital organs. We have developed an approach for ground truth generation and iterative training of 3D nuclear segmentation models, which we applied to popular CellPose, PlantSeg and StarDist algorithms. We provide two high-quality models trained on plant nuclei that enable 3D segmentation of nuclei in datasets obtained from fixed or live samples, acquired from different plant and animal tissues, and stained with various nuclear stains or fluorescent protein-based nuclear reporters. We also share a diverse high-quality training dataset of about 10,000 nuclei. Furthermore, we advanced the MorphoGraphX analysis and visualization software by, among other things, providing a method for linking 3D segmented nuclei to their surrounding cells in 3D digital organs. We found that the nuclear-to-cell volume ratio varies between different ovule tissues and during the development of a tissue. Finally, we extended the PlantSeg 3D segmentation pipeline with a proofreading tool that uses 3D segmented nuclei as seeds to correct cell segmentation errors in difficult-to-segment tissues.

Navigating the data frontier in science assessment: Advancing data augmentation strategies for machine learning applications with generative artificial intelligence

Machine learning (ML) techniques are commonly seen as an inductive learning procedure, typically involving the identification of patterns in a specific training dataset to make predictions in novel contexts. By doing so, the performance and generalizability of these techniques often rely on the quality and quantity of the available training data. However, gathering a diverse training dataset that captures multiple nuances of students’ reasoning poses challenges in educational settings due to resource constraints. We compared three data augmentation strategies to address this issue: collecting additional student data, utilizing chatbots to paraphrase existing responses, and prompting chatbots to generate synthetic responses. We found that leveraging data augmentation significantly improved ML model performance. In detail, combining authentic and/or paraphrased responses with chatbot responses yielded the best machine-human score agreements across various validation conditions. This data augmentation allowed us to expand our applied scoring rubric by introducing a more detailed categorization that better captured the level of causality in undergraduate chemistry students’ reasoning about reaction mechanisms. Together, these findings highlight effective possibilities for augmenting the size and heterogeneity of the training data to improve ML model performance and generalizability, introduce a more fine-grained categorization, and reduce human effort in data collection. In the future, these benefits may enhance the scalability of formative assessments that adaptively support students’ reasoning in postsecondary chemistry classes.

Global road extraction using a pseudo-label guided framework: from benchmark dataset to cross-region semi-supervised learning

ABSTRACT Recent advancements in satellite remote sensing technology and computer vision have enabled rapid extraction of road networks from massive, Very High-Resolution (VHR) satellite imagery. However, current road extraction methods face the following limitations: 1) Insufficient availability of accurate and diverse training datasets for global-scale road extraction; 2) Costly and time-consuming manual labeling of millions of road samples; and 3) Limited generalization ability of deep learning models across diverse global contexts, resulting in better performance for regions well-represented in the training dataset, but worse performance when faced with domain gaps. To address these challenges, a semi-supervised framework was developed in this study, which includes a global-scale benchmark dataset – termed GlobalRoadSet (GRSet) – and a pseudo-label guided semi-supervised road extraction network – termed GlobalRoadNetSF (GRNetSF). The GRSet dataset was constructed using high-resolution satellite imagery and open-source crowdsourced OpenStreetMap (OSM) data. It comprises 47,210 samples collected from 121 capital cities across six populated continents. The GRNetSF trains the network by generating pseudo-labels for unlabeled images, combined with a few labeled samples from the target region. To enhance the quality of the pseudo-labels, strong data augmentation perturbation and auxiliary feature perturbation techniques are employed to ensure model prediction consistency. The proposed GRNetSF_GRSet framework was implemented in over 30 cities worldwide, where most of the Intersection-over-Union (IoU) values increased by more than 10%. This outcome confirms its strong generalization ability.

Application of a U-Net Neural Network to the Puccinia sorghi-Maize Pathosystem.

Computer vision approaches to analyze plant disease data can be both faster and more reliable than traditional, manual methods. However, the requirement of manually annotating training data for the majority of machine learning applications can present a challenge for pipeline development. Here, we describe a machine learning approach to quantify Puccinia sorghi incidence on maize leaves utilizing U-Net convolutional neural network models. We analyzed several U-Net models with increasing amounts of training image data, either randomly chosen from a large data pool or randomly chosen from a subset of disease time course data. As the training dataset size increases, the models perform better, but the rate of performance decreases. Additionally, the use of a diverse training dataset can improve model performance and reduce the amount of annotated training data required for satisfactory performance. Models with as few as 48 whole-leaf training images are able to replicate the ground truth results within our testing dataset. The final model utilizing our entire training dataset performs similarly to our ground truth data, with an intersection over union value of 0.5002 and an F1 score of 0.6669. This work illustrates the capacity of U-Nets to accurately answer real-world plant pathology questions related to quantification and estimation of plant disease symptoms. [Formula: see text] Copyright © 2024 The Author(s). This is an open access article distributed under the CC BY-NC-ND 4.0 International license.

Integration of the Butina algorithm and ensemble learning strategies for the advancement of a pharmacophore ligand-based model: an in silico investigation of apelin agonists.

Introduction: 3D pharmacophore models describe the ligand's chemical interactions in their bioactive conformation. They offer a simple but sophisticated approach to decipher the chemically encoded ligand information, making them a valuable tool in drug design. Methods: Our research summarized the key studies for applying 3D pharmacophore models in virtual screening for 6,944 compounds of APJ receptor agonists. Recent advances in clustering algorithms and ensemble methods have enabled classical pharmacophore modeling to evolve into more flexible and knowledge-driven techniques. Butina clustering categorizes molecules based on their structural similarity (indicated by the Tanimoto coefficient) to create a structurally diverse training dataset. The learning method combines various individual pharmacophore models into a set of pharmacophore models for pharmacophore space optimization in virtual screening. Results: This approach was evaluated on Apelin datasets and afforded good screening performance, as proven by Receiver Operating Characteristic (AUC score of 0.994 ± 0.007), enrichment factor of (EF1% of 50.07 ± 0.211), Güner-Henry score of 0.956 ± 0.015, and F-measure of 0.911 ± 0.031. Discussion: Although one of the high-scoring models achieved statistically superior results in each dataset (AUC of 0.82; an EF1% of 19.466; GH of 0.131 and F1-score of 0.071), the ensemble learning method including voting and stacking method balanced the shortcomings of each model and passed with close performance measures.

Abstract 4884: Evaluating the utility of in silico variant annotation tools for cancer driver detection

Abstract In cancer genomics, precise variant annotation is crucial for clinical decisions, drug development, and research. The burgeoning genomic data offers an opportunity to use data-driven approaches to generate knowledge that supports clinical decisions. Particularly machine learning (ML) and Deep learning (DL), are becoming essential, as their application is fast, scalable, simple to implement, and generates reproducible results. The methods ranging from simple sequence-based alignment scoring to advanced algorithms like Logistic regression, Support vector machine, and Recurrent neural networks, have been employed by multiple variant annotation tools. This study compares in-silico methods available in Ensembl's Variant Effect Predictor (VEP) using a test dataset with COSMIC annotations. ML/DL success relies on robust training sets with comprehensive genomic variants features, including effects on transcription/translation, genomic context, annotation resources, in silico pathogenicity predictions, and population allele frequency. The training set, composed of known benign or pathogenic variants, serves as a reference for these algorithms to classify new and unseen variants. Our analysis reveals a limited concordance between the prediction algorithms. Despite comparable true/false positives/negatives, discrepancies persist in variant classification. Certain algorithms exhibit a propensity to over or under-call deleterious mutations. Some demonstrate a tendency to classify random variants in non-cancer genes as deleterious. Challenges include the absence of consensus on informative features, diverse training datasets, and restriction to well annotated proteins/transcripts. Balancing the sensitivity and false positives in detecting cancer drivers is crucial. Integrating individual prediction scores with ML algorithms enhances tool performance but comes with risk of error propagation, and limited accuracy. The study emphasises the need for context-specific variant classification tools, as many variants' impacts are cancer-type specific, and some may drive disease synergistically. Existing tools, designed for a "one variant - one score approach," struggle to capture complex associations, especially those dependent on changes in the tumour microenvironment. Highlighting areas for improvement, the study addresses the "black box" problem in decision processes. While limited interpretability might not hinder practical applications, tools should evolve to assess more complex associations guided by biology. Formal consensus, reference training datasets, and standards are deemed essential for developing next-generation tools. The envisioned context-dependent tools aim to streamline feature complexity, thereby mitigating the black box problem and advancing the accuracy and interpretability of cancer variant annotation. Citation Format: Madhumita, Zbyslaw Sondka, Jon Teague. Evaluating the utility of in silico variant annotation tools for cancer driver detection [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4884.

Rogue AI: Cautionary Cases in Neuroradiology and What We Can Learn From Them.

Introduction In recent years, artificial intelligence (AI) in medical imaging has undergone unprecedented innovation and advancement, sparking a revolutionary transformation in healthcare. The field of radiology is particularly implicated, as clinical radiologists are expected to interpret an ever-increasing number of complex cases in record time. Machine learning software purchased by our institution is expected to help our radiologists come to a more prompt diagnosis by delivering point-of-care quantitative analysis of suspicious findings and streamlining clinical workflow. This paper explores AI's impact on neuroradiology, an area accounting for a substantial portion of recent radiology studies. We present a case series evaluating an AI software's performance in detecting neurovascular findings, highlighting five cases where AI interpretations differed from radiologists' assessments. Our study underscores common pitfalls of AI in the context of CT head angiograms, aiming to guide future AI algorithms. Methods We conducted a retrospective case series study at Stony Brook University Hospital, a large medical center in Stony Brook, New York, spanning from October 1, 2021 to December 31, 2021, analyzing 140 randomly sampled CT angiograms using AI software. This software assessed various neurovascular parameters, and AI findings were compared with neuroradiologists' interpretations. Five cases with divergent interpretations were selected for detailed analysis. Results Five representative cases in which AI findings were discordant with radiologists' interpretations are presented with diagnoses including diffuse anoxic ischemic injury, cortical laminar necrosis, colloid cyst, right superficial temporal artery-to-middle cerebral artery(STA-MCA) bypass, and subacute bilateral subdural hematomas. Discussion The errors identified in our case series expose AI's limitations in radiology. Our case series reveals that AI's incorrect interpretations can stem from complexities in pathology, challenges in distinguishing densities, inability to identify artifacts, identifying post-surgical changes in normal anatomy, sensitivity limitations, and insufficient pattern recognition. AI's potential for improvement lies in refining its algorithms to effectively recognize and differentiate pathologies. Incorporating more diverse training datasets, multimodal data, deep-reinforcement learning, clinical context, and real-time learning capabilities are some ways to improve AI's performance in the field of radiology. Conclusion Overall, it is apparent that AI applications in radiology have much room for improvement before becoming more widely integrated into clinical workflows. While AI demonstrates remarkable potential to aid in diagnosis and streamline workflows, our case series highlights common pitfalls that underscore the need for continuous improvement. By refining algorithms, incorporating diverse datasets, embracing multimodal information, and leveraging innovative machine learning strategies, AI's diagnostic accuracy can be significantly improved.

Improving Generalizability of PET DL Algorithms: List-Mode Reconstructions Improve DOTATATE PET Hepatic Lesion Detection Performance.

Deep learning (DL) algorithms used for DOTATATE PET lesion detection typically require large, well-annotated training datasets. These are difficult to obtain due to low incidence of gastroenteropancreatic neuroendocrine tumors (GEP-NETs) and the high cost of manual annotation. Furthermore, networks trained and tested with data acquired from site specific PET/CT instrumentation, acquisition and processing protocols have reduced performance when tested with offsite data. This lack of generalizability requires even larger, more diverse training datasets. The objective of this study is to investigate the feasibility of improving DL algorithm performance by better matching the background noise in training datasets to higher noise, out-of-domain testing datasets. 68Ga-DOTATATE PET/CT datasets were obtained from two scanners: Scanner1, a state-of-the-art digital PET/CT (GE DMI PET/CT; n = 83 subjects), and Scanner2, an older-generation analog PET/CT (GE STE; n = 123 subjects). Set1, the data set from Scanner1, was reconstructed with standard clinical parameters (5 min; Q.Clear) and list-mode reconstructions (VPFXS 2, 3, 4, and 5-min). Set2, data from Scanner2 representing out-of-domain clinical scans, used standard iterative reconstruction (5 min; OSEM). A deep neural network was trained with each dataset: Network1 for Scanner1 and Network2 for Scanner2. DL performance (Network1) was tested with out-of-domain test data (Set2). To evaluate the effect of training sample size, we tested DL model performance using a fraction (25%, 50% and 75%) of Set1 for training. Scanner1, list-mode 2-min reconstructed data demonstrated the most similar noise level compared that of Set2, resulting in the best performance (F1 = 0.713). This was not significantly different compared to the highest performance, upper-bound limit using in-domain training for Network2 (F1 = 0.755; p-value = 0.103). Regarding sample size, the F1 score significantly increased from 25% training data (F1 = 0.478) to 100% training data (F1 = 0.713; p < 0.001). List-mode data from modern PET scanners can be reconstructed to better match the noise properties of older scanners. Using existing data and their associated annotations dramatically reduces the cost and effort in generating these datasets and significantly improves the performance of existing DL algorithms. List-mode reconstructions can provide an efficient, low-cost method to improve DL algorithm generalizability.

SlumberNet: deep learning classification of sleep stages using residual neural networks

Sleep research is fundamental to understanding health and well-being, as proper sleep is essential for maintaining optimal physiological function. Here we present SlumberNet, a novel deep learning model based on residual network (ResNet) architecture, designed to classify sleep states in mice using electroencephalogram (EEG) and electromyogram (EMG) signals. Our model was trained and tested on data from mice undergoing baseline sleep, sleep deprivation, and recovery sleep, enabling it to handle a wide range of sleep conditions. Employing k-fold cross-validation and data augmentation techniques, SlumberNet achieved high levels of overall performance (accuracy = 97%; F1 score = 96%) in predicting sleep stages and showed robust performance even with a small and diverse training dataset. Comparison of SlumberNet's performance to manual sleep stage classification revealed a significant reduction in analysis time (~ 50 × faster), without sacrificing accuracy. Our study showcases the potential of deep learning to facilitate sleep research by providing a more efficient, accurate, and scalable method for sleep stage classification. Our work with SlumberNet further demonstrates the power of deep learning in mouse sleep research.