Generalizable Model Research Articles

User-friendly prognostic model for rectal neuroendocrine tumours: In the era of precision management

In this letter, we explore into the potential role of the recent study by Zeng et al . Rectal neuroendocrine tumours (rNETs) are rare, originate from peptidergic neurons and neuroendocrine cells, and express corresponding markers. Although most rNETs patients have a favourable prognosis, the median survival period significantly decreases when high-risk factors, such as larger tumours, poorer differentiation, and lymph node metastasis exist, are present. Clinical prediction models play a vital role in guiding diagnosis and prognosis in health care, but their complex calculation formulae limit clinical use. Moreover, the prognostic models that have been developed for rNETs to date still have several limitations, such as insufficient sample sizes and the lack of external validation. A high-quality prognostic model for rNETs would guide treatment and follow-up, enabling the precise formulation of individual patient treatment and follow-up plans. The future development of models for rNETs should involve closer collaboration with statistical experts, which would allow the construction of clinical prediction models to be standardized and robust, accurate, and highly generalizable prediction models to be created, ultimately achieving the goal of precision medicine.

Predicting Therapeutic Response to Hypoglossal Nerve Stimulation Using Deep Learning.

To develop and validate machine learning (ML) and deep learning (DL) models using drug-induced sleep endoscopy (DISE) images to predict the therapeutic efficacy of hypoglossal nerve stimulator (HGNS) implantation. Patients who underwent DISE and subsequent HGNS implantation at a tertiary care referral center were included. Six DL models and five ML algorithms were trained on images from the base of tongue (BOT) and velopharynx (VP) from patients classified as responders or non-responders as defined by Sher's criteria (50% reduction in apnea-hypopnea index (AHI) and AHI < 15 events/h). Precision, recall, F1 score, and overall accuracy were evaluated as measures of performance. In total, 25,040 images from 127 patients were included, of which 16,515 (69.3%) were from responders and 8,262 (30.7%) from non-responders. Models trained on the VP dataset had greater overall accuracy when compared to BOT alone and combined VP and BOT image sets, suggesting that VP images contain discriminative features for identifying therapeutic efficacy. The VCG-16 DL model had the best overall performance on the VP image set with high training accuracy (0.833), F1 score (0.78), and recall (0.883). Among ML models, the logistic regression model had the greatest accuracy (0.685) and F1 score (0.813). Deep neural networks have potential to predict HGNS therapeutic efficacy using images from DISE, facilitating better patient selection for implantation. Development of multi-institutional data and image sets will allow for development of generalizable predictive models. NA Laryngoscope, 134:5210-5216, 2024.

Triplet Adaptation Framework for Robust Semi-Supervised Learning.

Semi-supervised learning (SSL) suffers from severe performance degradation when labeled and unlabeled data come from inconsistent and imbalanced distribution. Nonetheless, there is a lack of theoretical guidance regarding a remedy for this issue. To bridge the gap between theoretical insights and practical solutions, we embark to an analysis of generalization bound of classic SSL algorithms. This analysis reveals that distribution inconsistency between unlabeled and labeled data can cause a significant generalization error bound. Motivated by this theoretical insight, we present a Triplet Adaptation Framework (TAF) to reduce the distribution divergence and improve the generalization of SSL models. TAF comprises three adapters: Balanced Residual Adapter, aiming to map the class distribution of labeled and unlabeled data to a uniform distribution for reducing class distribution divergence; Representation Adapter, aiming to map the representation distribution of unlabeled data to labeled one for reducing representation distribution divergence; and Pseudo-Label Adapter, aiming to align the predicted pseudo-labels with the class distribution of unlabeled data, thereby preventing erroneous pseudo-labels from exacerbating representation divergence. These three adapters collaborate synergistically to reduce the generalization bound, ultimately achieving a more robust and generalizable SSL model. Extensive experiments across various robust SSL scenarios validate the efficacy of our method.

When Invariant Representation Learning Meets Label Shift: Insufficiency and Theoretical Insights.

As a crucial step toward real-world learning scenarios with changing environments, dataset shift theory and invariant representation learning algorithm have been extensively studied to relax the identical distribution assumption in classical learning setting. Among the different assumptions on the essential of shifting distributions, generalized label shift (GLS) is the latest developed one which shows great potential to deal with the complex factors within the shift. In this paper, we aim to explore the limitations of current dataset shift theory and algorithm, and further provide new insights by presenting a comprehensive understanding of GLS. From theoretical aspect, two informative generalization bounds are derived, and the GLS learner are proved to be sufficiently close to optimal target model from the Bayesian perspective. The main results show the insufficiency of invariant representation learning, and prove the sufficiency and necessity of GLS correction for generalization, which provide theoretical supports and innovations for exploring generalizable model under dataset shift. From methodological aspect, we provide a unified view of existing shift correction frameworks, and propose a kernel embedding-based correction algorithm (KECA) to minimize the generalization error and achieve successful knowledge transfer. Both theoretical results and extensive experiment evaluations demonstrate the sufficiency and necessity of GLS correction for addressing dataset shift and the superiority of proposed algorithm.

Abstract 4124675: Deep Learning Screening of Cardiac MRIs Uncovers Undiagnosed Hypertrophic Cardiomyopathy in the UK BioBank

Introduction: The prevalence of hypertrophic cardiomyopathy (HCM) in the UK Biobank based on ICD-10 codes (.07%) is lower than global estimates of disease prevalence (0.2 - 0.5%). Prior studies using this data have remarked on the limitations of findings given likely underdiagnosis. The availability of cardiac MRI scans on a fraction of the participants offers an opportunity to identify missed diagnoses. Aims: This study seeks to utilize a generalizable deep learning model to detect likely cases of undiagnosed hypertrophic cardiomyopathy from cardiac MRIs in the UK Biobank. Methods: The foundational model was trained on a multi-institutional dataset of 14,073 cardiac MRIs via a self-supervised contrastive learning approach that sought to minimize the divergence between scans and their associated radiology reports. The pre-trained model was fine-tuned to diagnose hypertrophic cardiomyopathy on a distinct cohort of 4,870 MRIs with 368 cases of HCM, achieving an AUC of 0.94. The fine-tuned model was applied to the UK Biobank cardiac MRI dataset to ascertain predicted probabilities of HCM. Cases exceeding a threshold of 95% – correlating to the top 0.5% of cases (expected specificity of 97% and sensitivity of 60%) – were screened in for manual reading. In a blinded fashion, a board-certified radiologist was tasked with diagnosing HCM on a sample of cases composed of high and low predicted probabilities. Results: Of the 43,017 patients with cardiac MRIs, only 9 (.02%) had an ICD diagnosis of HCM. 266 cardiac MRIs were manually reviewed: 216 had greater than 95% predicted probability of HCM; 50 negative controls were randomly selected amongst cases with predicted probability less than 10%. The radiologist concurred with an HCM diagnosis for 115 cases (sensitivity 53%, specificity 98%), 112 of which were previously undiagnosed. The prevalence of hypertension and aortic stenosis did not significantly differ between the cohort of true positives (69.2%) and false positives (76.6%). The corrected prevalence of HCM in the UK BioBank MRI cohort is estimated at 0.28%. Conclusions: The findings of this study illustrate the remarkable ability of a generalizable deep learning model to detect undiagnosed cases of a rare disease process from cardiac MRIs. This is an important milestone that may allow for widespread screening of hypertrophic cardiomyopathy while minimizing demand for radiologist labor, and thereby allow patients to reap the substantial benefits of earlier treatment.

A Generalizable Causal-Invariance-Driven Segmentation Model for Peripancreatic Vessels.

Segmenting peripancreatic vessels in CT, including the superior mesenteric artery (SMA), the coeliac artery (CA), and the partial portal venous system (PPVS), is crucial for preoperative resectability analysis in pancreatic cancer. However, the clinical applicability of vessel segmentation methods is impeded by the low generalizability on multi-center data, mainly attributed to the wide variations in image appearance, namely the spurious correlation factor. Therefore, we propose a causal-invariance-driven generalizable segmentation model for peripancreatic vessels. It incorporates interventions at both image and feature levels to guide the model to capture causal information by enforcing consistency across datasets, thus enhancing the generalization performance. Specifically, firstly, a contrast-driven image intervention strategy is proposed to construct image-level interventions by generating images with various contrast-related appearances and seeking invariant causal features. Secondly, the feature intervention strategy is designed, where various patterns of feature bias across different centers are simulated to pursue invariant prediction. The proposed model achieved high DSC scores (79.69%, 82.62%, and 83.10%) for the three vessels on a cross-validation set containing 134 cases. Its generalizability was further confirmed on three independent test sets of 233 cases. Overall, the proposed method provides an accurate and generalizable segmentation model for peripancreatic vessels and offers a promising paradigm for increasing the generalizability of segmentation models from a causality perspective. Our source codes will be released at https://github.com/ SJTUBME-QianLab/PC_VesselSeg.

Combining transfer learning and statistical measures to predict performance of composite materials with limited data

AbstractPredicting the performance of composite materials is crucial for their application in civil infrastructure, yet limited experimental data often hinder the development of accurate and generalizable models. This study introduces a deep neural network (DNN) approach that combines summarizing statistics (SS) and transfer learning (TL)—termed the SSTL‐DNN approach—to address data scarcity in modeling composite materials. The computational novelty lies in the SS method's ability to extract comprehensive information from limited datasets by converting complex constitutive laws into concise statistical representations, thereby enabling efficient and effective model training. Simultaneously, the TL method enhances computational efficiency by leveraging knowledge from related tasks with abundant data to improve learning in the target task with scarce data. This combination not only reduces dependency on large datasets but also significantly improves model generalization. The proposed SSTL‐DNN approach is validated through two case studies: fiber‐reinforced polymer confined concrete and engineered cementitious composites. In both case studies, the SSTL‐DNN model reduces the required dataset size by up to 75% and decreases the validation error by 39%, compared to traditional deep learning models. These results demonstrate that the SSTL‐DNN approach not only overcomes data scarcity but also provides accurate predictions and generalization to unseen data, offering a practical solution for modeling composite materials with limited data.

Novel Approaches for the Early Detection of Glaucoma Using Artificial Intelligence.

If left untreated, glaucoma-the second most common cause of blindness worldwide-causes irreversible visual loss due to a gradual neurodegeneration of the retinal ganglion cells. Conventional techniques for identifying glaucoma, like optical coherence tomography (OCT) and visual field exams, are frequently laborious and dependent on subjective interpretation. Through the fast and accurate analysis of massive amounts of imaging data, artificial intelligence (AI), in particular machine learning (ML) and deep learning (DL), has emerged as a promising method to improve the early detection and management of glaucoma. The purpose of this study is to examine the current uses of AI in the early diagnosis, treatment, and detection of glaucoma while highlighting the advantages and drawbacks of different AI models and algorithms. In addition, it aims to determine how AI technologies might transform glaucoma treatment and suggest future lines of inquiry for this area of study. A thorough search of databases, including Web of Science, PubMed, and Scopus, was carried out to find pertinent papers released until August 2024. The inclusion criteria were limited to research published in English in peer-reviewed publications that used AI, ML, or DL to diagnose or treat glaucoma in human subjects. Articles were chosen and vetted according to their quality, contribution to the field, and relevancy. Convolutional neural networks (CNNs) and other deep learning algorithms are among the AI models included in this paper that have been shown to have excellent sensitivity and specificity in identifying glaucomatous alterations in fundus photos, OCT scans, and visual field tests. By automating standard screening procedures, these models have demonstrated promise in distinguishing between glaucomatous and healthy eyes, forecasting the course of the disease, and possibly lessening the workload of physicians. Nonetheless, several significant obstacles remain, such as the requirement for various training datasets, outside validation, decision-making transparency, and handling moral and legal issues. Artificial intelligence (AI) holds great promise for improving the diagnosis and treatment of glaucoma by facilitating prompt and precise interpretation of imaging data and assisting in clinical decision making. To guarantee wider accessibility and better patient results, future research should create strong generalizable AI models validated in various populations, address ethical and legal matters, and incorporate AI into clinical practice.

ECGWiz: a unified and generalizable AI model for comprehensive ECG diagnosis

Abstract Background Deep learning has demonstrated diagnostic performance in electrocardiogram (ECG) analysis comparable to that of clinicians. However, existing models primarily excel in diagnosing common conditions such as arrhythmias and heart blocks, which are easily identifiable using standard predictor variables. Conversely, models aiming to uncover novel insights, such as detecting structural changes or the presence of myocardial scar, often lack generalizability due to insufficient data, model architecture limitations, and other factors. Moreover, these models frequently encounter challenges, such as a decline in performance when applied to external datasets. To address these issues comprehensively, we propose the ECGWiz — a unified model capable of tackling diverse ECG-related problems with consistently high performance. Objective To develop a generalizable Foundation AI model for ECG diagnosis capable of detecting diverse cardiac abnormalities and diseases using few training samples. Methods We utilized publicly available datasets like PTB-XL, CPSC, etc. To build our model, totaling approximately 80,000 ECGs. These 12 lead ECGs are scaled and filtered before using them for training. The AI model is pretrained using Self Supervised Learning and Contrastive Learning strategies. Further, the model undergoes supervised training using a hierarchy of labels that increase in granularity, encompassing arrhythmias, heart blocks, other cardiac abnormalities, heart failure, sudden cardiac death and more. Lastly, the model is prepared for Few-Shot Learning — where the model learns from a minimal set of samples (5 to 10 ECGs) of the target disease. For testing the model, we are using unique datasets like PROSE which consists of non-ischemic cardiomyopathy patients with and without myocardial scaring detected using cardiac MRI imaging at the Johns Hopkins University. Results Our model, trained on a dataset of only 5 ECG recordings, demonstrated a 74% accuracy in identifying myocardial scarring as detected by LGE MRI, utilizing a standard 12-lead ECG. This performance is comparable to that of conventional deep learning methods, which necessitate datasets containing upwards of 120 ECG recordings. Conclusion Our model represents a promising stride in creating a unified and adaptable foundation AI model for comprehensive ECG-based diagnosis. Its capacity to learn from a minimal set of examples signifies a robust framework that can evolve and adapt to new challenges in ECG diagnosis. Ongoing improvements, particularly through dataset expansion, underscore its potential to enhance cardiovascular diagnostics, offering precise and efficient patient care.

Subject-aware PET Denoising with Contrastive Adversarial Domain Generalization.

Recent advances in deep learning (DL) have greatly improved the performance of positron emission tomography (PET) denoising performance. However, DL model performance can vary a lot across subjects, due to the large variability of the count levels and spatial distributions. A generalizable DL model that mitigates the subject-wise variations is highly expected toward a reliable and trustworthy system for clinical application. In this work, we propose a contrastive adversarial learning framework for subject-wise domain generalization (DG). Specifically, we configure a contrastive discriminator in addition to the UNet-based denoising module to check the subject-related information in the bottleneck feature, while the denoising module is adversarially trained to enforce the extraction of subject-invariant features. The sampled low-count realizations from the list-mode data are used as anchor-positive pairs to be close to each other, while the other subjects are used as negative samples to be distributed far away. We evaluated on 97 18F-MK6240 tau PET studies, each having 20 noise realizations with 25% fractions of events. Training, validation, and testing were implemented using 1400, 120, and 420 pairs of 3D image volumes in a subject-independent manner. The proposed contrastive adversarial DG demonstrated superior denoising performance than conventional UNet without subject-wise DG and cross-entropy-based adversarial DG.

Multi-scale retrofit pathways for improving building performance and energy equity across cities: A UBEM framework

Institutional inequality has created an environment in which our physical surroundings reflect racial and economic biases – causing inequitable thermal comfort and energy consumption across neighborhoods. Just as energy and thermal comfort challenges are unique to each climate and built environment, the solutions to address these issues should be optimized for their context. We introduce a novel integrated data-driven and building energy simulation method that incorporates urban context, satellite imagery, and socioeconomic data to analyze multi-scale retrofit scenarios across disparate communities. We apply this method to case studies in New York City and Los Angeles to analyze energy and thermal comfort inequities across distinct climate zones and urban morphologies. Through parametric analysis, we demonstrate the efficacy of multi-scale retrofits at improving building performance and reducing existing thermal inequities. Our research shows that retrofit effectiveness is influenced by the microclimatic changes induced by the urban context and existing infrastructural inequity. For example, instantaneous building cooling energy demand in a disadvantaged community of Los Angeles is reduced more due to window retrofits (around 30 kWh) than by a greenspace retrofit (around 3 kWh). Greenspace installation dissipates more heat overnight and in the early morning, whereas window retrofits reduce solar heat gain during the day, showing the different heat reduction pathways that retrofits can achieve. Although the window retrofits lead to an order of magnitude higher utility cost savings ($70 savings from windows vs $9 for greenspace and $12 for reduced cars), the district-scale retrofits show promising pathways for reducing the cost burden across all residents. We also analyze pathways for achieving city-specific CO2 reduction targets and find that building-level retrofits are most effective at meeting building performance standards (reducing CO2 by 40%), whereas district-level retrofits should be used to achieve city-scale climate goals because of the sequestered and abated emissions offered by these pathways. This generalizable modeling framework empowers policymakers, urban planners, and building owners around the world to analyze pathways for meeting their climate action plan goals by reducing extreme heat, reducing greenhouse emissions, and reducing energy cost burden while improving environmental justice in their cities.

Soil density specification ranges to optimise plant water use and root growth for phytocapping and urban greening projects

Soil density or compaction is one of the key variables that need to be specified for designed soil profiles as part of urban greening. Examples include land rehabilitation (mining), waste containment (phytocapping), stormwater infrastructure (bioretention basins) and green infrastructure (green roofs and street trees). This study investigates the impact of a range of specified soil densities on plant water use and root development. Native Australian plant species selected for the study include: the C4 and C3 grasses, Themeda triandra and Microlaena stipoides respectively; the Eucalyptus trees, E. camaldulensis and E. cladocalyx; and the nitrogen-fixing pioneer trees, Acacia mearnsii and Allocasuarina verticillata. The plants were established at four soil density (compaction) levels: 72, 77, 82, and 87 % MDD (maximum dry density) in tall cylinders with weekly plant water use measurement over 8 months. Root growth was analysed using WinRhizo image analysis. The experimental results were used to create generalisable models for root length density (RLD), root diameter and plant water use. The models for RLD and plant water use were parabolic in nature, revealing clear optimum ranges that could be used to guide soil density specification. Root diameter provided additional insight into the allocation of resources to root thickening above a threshold soil density of 87 % MDD, indicating plant allocation of resources towards penetrating highly compacted soils. There were correlations between plant water use and RLD that were moderate and significant, particularly for grasses. Notably, T. triandra had the greatest mean RLD, thickest roots and plant water use at 16.8cm/cm3, 0.18 mm and 10mm/week, respectively. Findings demonstrate that RLD and plant water use can be optimised together within practically achievable soil density specification ranges that are sensitive to the pitfalls of both excessively low and excessively high soil densities. Recommended soil density specification ranges include: 75–82 % MDD with plant performance within 5 % of optimum (considered excellent), 74–84 % MDD with plant performance within 10 % of optimum (considered good) and 73–85 % MDD with plant performance within 15 % of optimum (considered fair). Within these ranges, plant water use and root growth performance is well balanced with practical achievement of the soil density ranges. Due to the use of %MDD, the modelled results can be usefully generalised for any soil type. Implementation of these specifications for urban greening and phytocap projects will optimise plant growth, transpiration and hydrological function while maintaining root networks essential for establishing and maintaining resilient living infrastructure.

Inferring post-transcriptional regulation within and across cell types in human testis.

Single-cell tissue atlases commonly use RNA abundances as surrogates for protein abundances. Yet, protein abundance also depends on the regulation of protein synthesis and degradation rates. To estimate the contributions of such post transcriptional regulation, we quantified the proteomes of 5,883 single cells from human testis using 3 distinct mass spectrometry methods (SCoPE2, pSCoPE, and plexDIA). To distinguish between biological and technical factors contributing to differences between protein and RNA levels, we developed BayesPG , a Bayesian model of transcript and protein abundance that systematically accounts for technical variation and infers biological differences. We use BayesPG to jointly model RNA and protein data collected from 29,709 single cells across different methods and datasets. BayesPG estimated consensus mRNA and protein levels for 3,861 gene products and quantified the relative protein-to-mRNA ratio (rPTR) for each gene across six distinct cell types in samples from human testis. About 28% of the gene products exhibited significant differences at protein and RNA levels and contributed to about 1, 500 significant GO groups. We observe that specialized and context specific functions, such as those related to spermatogenesis are regulated after transcription. Among hundreds of detected post translationally modified peptides, many show significant abundance differences across cell types. Furthermore, some phosphorylated peptides covary with kinases in a cell-type dependent manner, suggesting cell-type specific regulation. Our results demonstrate the potential of inferring protein regulation in from single-cell proteogenomic data and provide a generalizable model, BayesPG , for performing such analyses.

RFMiD: Retinal Image Analysis for multi-Disease Detection challenge

In the last decades, many publicly available large fundus image datasets have been collected for diabetic retinopathy, glaucoma, and age-related macular degeneration, and a few other frequent pathologies. These publicly available datasets were used to develop a computer-aided disease diagnosis system by training deep learning models to detect these frequent pathologies. One challenge limiting the adoption of a such system by the ophthalmologist is, computer-aided disease diagnosis system ignores sight-threatening rare pathologies such as central retinal artery occlusion or anterior ischemic optic neuropathy and others that ophthalmologists currently detect. Aiming to advance the state-of-the-art in automatic ocular disease classification of frequent diseases along with the rare pathologies, a grand challenge on “Retinal Image Analysis for multi-Disease Detection” was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI - 2021). This paper, reports the challenge organization, dataset, top-performing participants solutions, evaluation measures, and results based on a new “Retinal Fundus Multi-disease Image Dataset” (RFMiD). There were two principal sub-challenges: disease screening (i.e. presence versus absence of pathology — a binary classification problem) and disease/pathology classification (a 28-class multi-label classification problem). It received a positive response from the scientific community with 74 submissions by individuals/teams that effectively entered in this challenge. The top-performing methodologies utilized a blend of data-preprocessing, data augmentation, pre-trained model, and model ensembling. This multi-disease (frequent and rare pathologies) detection will enable the development of generalizable models for screening the retina, unlike the previous efforts that focused on the detection of specific diseases.

A generalizable and interpretable model for early warning of pest-induced crop diseases using environmental data

Pest infestations and resulting crop diseases threaten global food security. Traditional pest and disease monitoring methods are time-consuming and prone to delays, thus necessitating the development of effective prediction strategies that facilitate early and timely detection. In response to this challenge, this research proposes a generalizable and interpretable machine learning model to predict two major rice pests—Green Leafhopper and Yellow Stem Borer—and a No Pest class using environmental data collected from various regions in India. The dataset, comprising factors such as temperature, humidity, and rainfall, underwent rigorous preprocessing. Key features like temperature difference, humidity difference, and vapor pressure deficit were engineered to enhance the model’s performance. Twelve baseline models were trained and their performance was evaluated using F1 scores and AUC values due to the imbalanced nature of the dataset. Through statistical analysis of baseline models, Random Forest, Balanced Random Forest, XGBoost, and CatBoost models, were selected for hyperparameter tuning via Optuna. The tuned CatBoost model demonstrated superior performance, achieving AUC values of 0.99 for Green Leafhoppers, 0.98 for Yellow Stem Borers, 1.00 for No Pest class and an overall F1 score of 0.9414 with a mean AUC value of 0.9912 across all classes. Additionally, Explainable Artificial Intelligence techniques, particularly SHAP, were employed to interpret the model’s predictions, identifying the relative importance of environmental factors in pest occurrence. This interpretability aligns the model’s predictions with established agronomic knowledge, enhancing its practical utility for early pest and disease detection. The generalizability of the proposed model suggests it can be adapted to other crops and regions, offering a valuable tool for early warning systems in agriculture, promoting sustainable practices, and reducing crop losses due to pests and diseases.

BrainQCNet: A Deep Learning attention-based model for the automated detection of artifacts in brain structural MRI scans

Abstract Analyses of structural MRI (sMRI) data depend on robust upstream data quality control (QC). It is also crucial that researchers seek to retain maximal amounts of data to ensure reproducible, generalizable models and to avoid wasted effort, including that of participants. The time-consuming and difficult task of manual QC evaluation has prompted the development of tools for the automatic assessment of brain sMRI scans. Existing tools have proved particularly valuable in this age of Big Data; as datasets continue to grow, reducing execution time for QC evaluation will be of considerable benefit. The development of Deep Learning (DL) models for artifact detection in structural MRI scans offers a promising avenue toward fast, accurate QC evaluation. In this study, we trained an interpretable Deep Learning model, ProtoPNet, to classify minimally preprocessed 2D slices of scans that had been manually annotated with a refined quality assessment (ABIDE 1; n = 980 scans). To evaluate the best model, we applied it to 2141 ABCD T1-weighted MRI scans for which gold-standard manual QC annotations were available. We obtained excellent accuracy: 82.4% for good quality scans (Pass), 91.4% for medium to low quality scans (Fail). Further validation using 799 T1w MRI scans from ABIDE 2 and 750 T1w MRI scans from ADHD-200 confirmed the reliability of our model. Accuracy was comparable to or exceeded that of existing ML models, with fast processing and prediction time (1 minute per scan, GPU machine, CUDA-compatible). Our attention model also performs better than traditional DL (i.e., convolutional neural network models) in detecting poor quality scans. To facilitate faster and more accurate QC prediction for the neuroimaging community, we have shared the model that returned the most reliable global quality scores as a BIDS-app (https://github.com/garciaml/BrainQCNet).

Artificial Intelligence for Prediction of Shunt Response in Idiopathic Normal Pressure Hydrocephalus: A Systematic Review

BackgroundIdiopathic normal pressure hydrocephalus (iNPH) is a reversible cause of dementia, typically treated with shunt surgery, although outcomes vary. Artificial intelligence (AI) advancements could improve predictions of shunt response (SR) by analyzing extensive data sets. MethodsWe conducted a systematic review to assess AI's effectiveness in predicting SR in iNPH. Studies using AI or machine learning (ML) algorithms for SR prediction were identified through searches in MEDLINE, EMBASE, and Web of Science up to September 2023, adhering to Synthesis Without Meta-Analysis reporting guidelines. ResultsOut of 3541 studies identified, 33 were assessed for eligibility, and 8 involving 479 patients were included. Study sample sizes varied from 28 to 132 patients. Common data inputs included imaging/radiomics (62.5%) and demographics (37.5%), with Support Vector Machine being the most frequently used ML algorithm (87.5%). Two studies compared multiple algorithms. Only four studies reported the Area Under the Curve (AUC) values, which ranged between 0.80 and 0.94. The results highlighted inconsistency in outcome measures, data heterogeneity, and potential biases in the models used. ConclusionsWhile AI shows promise for improving iNPH management, there is a need for standardized data and extensive validation of AI models to enhance their clinical utility. Future research should aim to develop robust and generalizable AI models for more effective diagnosis and management of iNPH.

Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert

In response to real-world scenarios, the domain generalization (DG) problem has spurred considerable research in person re-identification (ReID). This challenge arises when the target domain, which is significantly different from the source domains, remains unknown. However, the performance of current DG ReID relies heavily on labor-intensive source domain annotations. Considering the potential of unlabeled data, we investigate unsupervised domain generalization (UDG) in ReID. Our goal is to create a model that can generalize from unlabeled source domains to semantically retrieve images in an unseen target domain. To address this, we propose a new approach that trains a domain-agnostic expert (DaE) for unsupervised domain-generalizable person ReID. This involves independently training multiple experts to account for label space inconsistencies between source domains. At the same time, the DaE captures domain-generalizable information for testing. Our experiments demonstrate the effectiveness of this method for learning generalizable features under the UDG setting. The results demonstrate the superiority of our method over state-of-the-art techniques. We will make our code and models available for public use.

An adaptive learning strategy for surrogate modeling of high-dimensional functions - Application to unsteady hypersonic flows in chemical nonequilibrium

Many engineering applications rely on the evaluation of expensive, non-linear high-dimensional functions. In this paper, we propose the RONAALP algorithm (Reduced Order Nonlinear Approximation with Adaptive Learning Procedure) to incrementally learn, on-the-fly as the application progresses, a fast and accurate reduced-order surrogate model of a target function. First, a combination of nonlinear auto-encoder, community clustering, and radial basis function networks allows us to learn an efficient and compact surrogate model with limited training data. Secondly, an active learning procedure overcomes any extrapolation issues during the online stage by adapting the surrogate model with high-fidelity evaluations that fall outside its current validity range. This approach results in generalizable, fast, and accurate reduced-order models of high-dimensional functions. The method is demonstrated on three direct numerical simulations of hypersonic flows in chemical nonequilibrium. Accurate simulations of these flows rely on detailed thermochemical gas models that dramatically increase the cost of such calculations. Using RONAALP to learn a reduced-order thermodynamic model surrogate on-the-fly, the cost of such simulations was reduced by up to 75% while maintaining an error of less than 10% on relevant quantities of interest.

Fast Selection of Indoor Wireless Transmitter Locations With Generalizable Neural Network Propagation Models

Fast Selection of Indoor Wireless Transmitter Locations With Generalizable Neural Network Propagation Models