Generic Deep Learning Research Articles

An interpretable algorithm unrolling network inspired by general convolutional sparse coding for intelligent fault diagnosis of machinery

When applied in Intelligent Fault Diagnosis (IFD) of machinery, most of generic deep learning models lack interpretability in architecture design. From a signal processing perspective, we unroll the Alternating Direction Method of Multipliers (ADMM), the iterative algorithm for a general convolutional sparse coding denoising problem, into a deep neural network called the General ADMM Network (GADMM-Net) using algorithm unrolling. GADMM-Net has an encoder-decoder architecture where the encoder extracts signal features (i.e., sparse coefficients) and the decoder reconstructs denoised signals. As the encoder-decoder architecture is derived from the denoising algorithm, GADMM-Net is interpretable in its backbone architecture and inherits the prior domain knowledge behind the signal denoising problem. Compared with conventional and state-of-the-art IFD models, GADMM-Net performs more excellently in diagnostic accuracy, number of parameters, training set size requirement, and noise robustness. In the IFD task across different operating conditions, GADMM-Net shows domain generalization capabilities, without using any transfer learning strategy.

Targeting protein-ligand neosurfaces with a generalizable deep learning tool.

Molecular recognition events between proteins drive biological processes in living systems1. However, higher levels of mechanistic regulation have emerged, in which protein-protein interactions are conditioned to small molecules2-5. Despite recent advances, computational tools for the design of new chemically induced protein interactions have remained a challenging task for the field6,7. Here we present a computational strategy for the design of proteins that target neosurfaces, that is, surfaces arising from protein-ligand complexes. To develop this strategy, we leveraged a geometric deep learning approach based on learned molecular surface representations8,9 and experimentally validated binders against three drug-bound protein complexes: Bcl2-venetoclax, DB3-progesterone and PDF1-actinonin. All binders demonstrated high affinities and accurate specificities, as assessed by mutational and structural characterization. Remarkably, surface fingerprints previously trained only on proteins could be applied to neosurfaces induced by interactions with small molecules, providing a powerful demonstration of generalizability that is uncommon in other deep learning approaches. We anticipate that such designed chemically induced protein interactions will have the potential to expand the sensing repertoire and the assembly of new synthetic pathways in engineered cells for innovative drug-controlled cell-based therapies10.

MeTa Learning-Based Optimization of Unsupervised Domain Adaptation Deep Networks

This paper introduces a novel unsupervised domain adaptation (UDA) method, MeTa Discriminative Class-Wise MMD (MCWMMD), which combines meta-learning with a Class-Wise Maximum Mean Discrepancy (MMD) approach to enhance domain adaptation. Traditional MMD methods align overall distributions but struggle with class-wise alignment, reducing feature distinguishability. MCWMMD incorporates a meta-module to dynamically learn a deep kernel for MMD, improving alignment accuracy and model adaptability. This meta-learning technique enhances the model’s ability to generalize across tasks by ensuring domain-invariant and class-discriminative feature representations. Despite the complexity of the method, including the need for meta-module training, it presents a significant advancement in UDA. Future work will explore scalability in diverse real-world scenarios and further optimize the meta-learning framework. MCWMMD offers a promising solution to the persistent challenge of domain adaptation, paving the way for more adaptable and generalizable deep learning models.

A Generic Deep Learning–Based Computing Algorithm in Support of the Development of Instrumented Bikes

A Generic Deep Learning–Based Computing Algorithm in Support of the Development of Instrumented Bikes

Improved Generalizability in Medical Computer Vision: Hyperbolic Deep Learning in Multi-Modality Neuroimaging.

Deep learning has shown significant value in automating radiological diagnostics but can be limited by a lack of generalizability to external datasets. Leveraging the geometric principles of non-Euclidean space, certain geometric deep learning approaches may offer an alternative means of improving model generalizability. This study investigates the potential advantages of hyperbolic convolutional neural networks (HCNNs) over traditional convolutional neural networks (CNNs) in neuroimaging tasks. We conducted a comparative analysis of HCNNs and CNNs across various medical imaging modalities and diseases, with a focus on a compiled multi-modality neuroimaging dataset. The models were assessed for their performance parity, robustness to adversarial attacks, semantic organization of embedding spaces, and generalizability. Zero-shot evaluations were also performed with ischemic stroke non-contrast CT images. HCNNs matched CNNs' performance in less complex settings and demonstrated superior semantic organization and robustness to adversarial attacks. While HCNNs equaled CNNs in out-of-sample datasets identifying Alzheimer's disease, in zero-shot evaluations, HCNNs outperformed CNNs and radiologists. HCNNs deliver enhanced robustness and organization in neuroimaging data. This likely underlies why, while HCNNs perform similarly to CNNs with respect to in-sample tasks, they confer improved generalizability. Nevertheless, HCNNs encounter efficiency and performance challenges with larger, complex datasets. These limitations underline the need for further optimization of HCNN architectures. HCNNs present promising improvements in generalizability and resilience for medical imaging applications, particularly in neuroimaging. Despite facing challenges with larger datasets, HCNNs enhance performance under adversarial conditions and offer better semantic organization, suggesting valuable potential in generalizable deep learning models in medical imaging and neuroimaging diagnostics.

Retinal Image Augmentation using Composed GANs

Medical image analysis faces a significant challenge in the scarcity of annotated data, which is crucial for developing generalizable Deep Learning (DL) models that require extensive training data. Consequently, the field of medical image generation has garnered substantial interest and potential for further exploration. Besides widely employed data augmentation techniques, such as rotation, reflection, and scaling, Generative Adversarial Networks (GANs) have demonstrated the ability to effectively leverage additional information from datasets by generating synthetic samples from real images. In the context of retinal image synthesis, an image-to-image translation approach is frequently adopted to generate retinal images from available vessel maps, which can be scarce and resource-intensive to obtain. Deviating from prior work reliant on pre-existing vessel maps, this study proposes a learning-based model that is independent of vessel maps, utilizing Progressive Growing GAN (PGGAN) to generate vascular networks from random noise. The visual and quantitative evaluations conducted suggest that the majority of the images generated by the proposed model are substantially distinct from the training set while maintaining a high proportion of true image quality, underscoring the model's potential as a powerful tool for data augmentation.

Generative modeling of the Circle of Willis using 3D-StyleGAN

The circle of Willis (CoW) is a network of cerebral arteries with significant inter-individual anatomical variations. Deep learning has been used to characterize and quantify the status of the CoW in various applications for the diagnosis and treatment of cerebrovascular disease. In medical imaging, the performance of deep learning models is limited by the diversity and size of training datasets. To address medical data scarcity, generative AI models have been applied to generate synthetic vessel neuroimaging data. However, the proposed methods produce synthetic data with limited anatomical fidelity or downstream utility in tasks concerning vessel characteristics.We adapted the StyleGANv2 architecture to 3D to synthesize Time-of-Flight Magnetic Resonance Angiography (TOF MRA) volumes of the CoW. For generative modeling, we used 1782 individual TOF MRA scans from 6 open source datasets. To train the adapted 3D StyleGAN model with limited data we employed differentiable data augmentations, used mixed precision and a cropped region of interest of size 32 × 128 × 128 to tackle computational constraints. The performance was evaluated quantitatively using the Fréchet Inception Distance (FID), MedicalNet distance (MD) and Area Under the Curve of the Precision and Recall Curve for Distributions (AUC-PRD). Qualitative analysis was performed via a visual Turing test. We demonstrated the utility of generated data in a downstream task of multiclass semantic segmentation of CoW arteries. Vessel segmentation performance was assessed quantitatively using the Dice coefficient and the Hausdorff distance.The best-performing 3D StyleGANv2 architecture generated high-quality and diverse synthetic TOF MRA volumes (FID: 12.17, MD: 0.00078, AUC-PRD: 0.9610). Multiclass vessel segmentation models trained on synthetic data alone achieved comparable performance to models trained using real data in most arteries. The addition of synthetic data to a baseline training set improved segmentation performance in underrepresented artery segments, similar to the addition of real data.In conclusion, generative modeling of the Circle of Willis via synthesis of 3D TOF MRA data paves the way for generalizable deep learning applications in cerebrovascular disease. In the future, the extensions of the provided methodology to other medical imaging problems or modalities with the inclusion of pathological datasets has the potential to advance the development of more robust AI models for clinical applications.

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.

Simulating the Cellular Context in Synthetic Datasets for Cryo-Electron Tomography.

Cryo-electron tomography (cryo-ET) allows to visualize the cellular context at macromolecular level. To date, the impossibility of obtaining a reliable ground truth is limiting the application of deep learning-based image processing algorithms in this field. As a consequence, there is a growing demand of realistic synthetic datasets for training deep learning algorithms. In addition, besides assisting the acquisition and interpretation of experimental data, synthetic tomograms are used as reference models for cellular organization analysis from cellular tomograms. Current simulators in cryo-ET focus on reproducing distortions from image acquisition and tomogram reconstruction, however, they can not generate many of the low order features present in cellular tomograms. Here we propose several geometric and organization models to simulate low order cellular structures imaged by cryo-ET. Specifically, clusters of any known cytosolic or membrane-bound macromolecules, membranes with different geometries as well as different filamentous structures such as microtubules or actin-like networks. Moreover, we use parametrizable stochastic models to generate a high diversity of geometries and organizations to simulate representative and generalized datasets, including very crowded environments like those observed in native cells. These models have been implemented in a multiplatform open-source Python package, including scripts to generate cryo-tomograms with adjustable sizes and resolutions. In addition, these scripts provide also distortion-free density maps besides the ground truth in different file formats for efficient access and advanced visualization. We show that such a realistic synthetic dataset can be readily used to train generalizable deep learning algorithms.

A Pilot Study of Deep Learning-Based Monocular Depth Estimation from Fundus Photographs

The purpose of this study was to evaluate the feasibility of a generalizable deep-learning (DL) based system with no a priori knowledge of fundus photographs to generate monocular depth map information about optic disc structures from this imaging modality. Images of 30 stereo pairs of fundus photographs centered on the optic disc of 30 subjects were analyzed with this DL system to generate monocular depth maps using zero-shot cross-dataset transfer. These maps were registered onto reference standard depth maps derived from Optical Coherence Tomography. Accuracy of the DL system was assessed by the root of mean squared error (RMSE) between the estimate and reference standard. 47% of the total images from the dataset were successfully processed, with mean RMSE of 0.081. Our findings demonstrate that single image, monocular depth estimation with a generalizable DL system using zero-shot cross-dataset transfer applied to retinal color fundus photographs is feasible and has potential.

Privacy enhancing and generalizable deep learning with synthetic data for mediastinal neoplasm diagnosis

The success of deep learning (DL) relies heavily on training data from which DL models encapsulate information. Consequently, the development and deployment of DL models expose data to potential privacy breaches, which are particularly critical in data-sensitive contexts like medicine. We propose a new technique named DiffGuard that generates realistic and diverse synthetic medical images with annotations, even indistinguishable for experts, to replace real data for DL model training, which cuts off their direct connection and enhances privacy safety. We demonstrate that DiffGuard enhances privacy safety with much less data leakage and better resistance against privacy attacks on data and model. It also improves the accuracy and generalizability of DL models for segmentation and classification of mediastinal neoplasms in multi-center evaluation. We expect that our solution would enlighten the road to privacy-preserving DL for precision medicine, promote data and model sharing, and inspire more innovation on artificial-intelligence-generated-content technologies for medicine.

Training robust T1-weighted magnetic resonance imaging liver segmentation models using ensembles of datasets with different contrast protocols and liver disease etiologies

Image segmentation of the liver is an important step in treatment planning for liver cancer. However, manual segmentation at a large scale is not practical, leading to increasing reliance on deep learning models to automatically segment the liver. This manuscript develops a generalizable deep learning model to segment the liver on T1-weighted MR images. In particular, three distinct deep learning architectures (nnUNet, PocketNet, Swin UNETR) were considered using data gathered from six geographically different institutions. A total of 819 T1-weighted MR images were gathered from both public and internal sources. Our experiments compared each architecture’s testing performance when trained both intra-institutionally and inter-institutionally. Models trained using nnUNet and its PocketNet variant achieved mean Dice-Sorensen similarity coefficients>0.9 on both intra- and inter-institutional test set data. The performance of these models suggests that nnUNet and PocketNet liver segmentation models trained on a large and diverse collection of T1-weighted MR images would on average achieve good intra-institutional segmentation performance.

A light-weight and generalizable deep learning model for the prediction of COVID-19 from chest X-ray images

The detection of coronavirus disease (COVID-19) using standard laboratory tests, such as reverse transcription polymerase chain reaction (RT-PCR), is time-consuming. Complex medical imaging problems are currently being solved using machine learning and deep learning techniques. Our proposed solution utilizes chest radiography imaging techniques, which have shown to be a faster alternative for detecting COVID-19. We present an efficient and lightweight deep learning architecture for identifying COVID-19 using chest X-ray images which achieve 99.81% accuracy in intra-database testing and 100% accuracy in cross-validation testing on a separate data set. The results demonstrate the potential of our proposed model as a reliable tool for COVID-19 diagnosis using chest X-ray images, which can have a significant impact on improving the efficiency of COVID-19 diagnosis and treatment.

A generic deep learning architecture optimization method for edge device based on start-up latency reduction

In the promising Artificial Intelligence of Things technology, deep learning algorithms are implemented on edge devices to process data locally. However, high-performance deep learning algorithms are accompanied by increased computation and parameter storage costs, leading to difficulties in implementing huge deep learning algorithms on memory and power constrained edge devices, such as smartphones and drones. Thus various compression methods are proposed, such as channel pruning. According to the analysis of low-level operations on edge devices, existing channel pruning methods have limited effect on latency optimization. Due to data processing operations, the pruned residual blocks still result in significant latency, which hinders real-time processing of CNNs on edge devices. Hence, we propose a generic deep learning architecture optimization method to achieve further acceleration on edge devices. The network is optimized in two stages, Global Constraint and Start-up Latency Reduction, and pruning of both channels and residual blocks is achieved. Optimized networks are evaluated on desktop CPU, FPGA, ARM CPU, and PULP platforms. The experimental results show that the latency is reduced by up to 70.40%, which is 13.63% higher than only applying channel pruning and achieving real-time processing in the edge device.

Deep learning survival model predicts outcome after intracerebral hemorrhage from initial CT scan.

Predicting functional impairment after intracerebral hemorrhage (ICH) provides valuable information for planning of patient care and rehabilitation strategies. Current prognostic tools are limited in making long term predictions and require multiple expert-defined inputs and interpretation that make their clinical implementation challenging. This study aimed to predict long term functional impairment of ICH patients from admission non-contrast CT scans, leveraging deep learning models in a survival analysis framework. We used the admission non-contrast CT scans from 882 patients from the Massachusetts General Hospital ICH Study for training, hyperparameter optimization, and model selection, and 146 patients from the Yale New Haven ICH Study for external validation of a deep learning model predicting functional outcome. Disability (modified Rankin scale [mRS] > 2), severe disability (mRS > 4), and dependent living status were assessed via telephone interviews after 6, 12, and 24 months. The prediction methods were evaluated by the c-index and compared with ICH score and FUNC score. Using non-contrast CT, our deep learning model achieved higher prediction accuracy of post-ICH dependent living, disability, and severe disability by 6, 12, and 24 months (c-index 0.742 [95% CI -0.700 to 0.778], 0.712 [95% CI -0.674 to 0.752], 0.779 [95% CI -0.733 to 0.832] respectively) compared with the ICH score (c-index 0.673 [95% CI -0.662 to 0.688], 0.647 [95% CI -0.637 to 0.661] and 0.697 [95% CI -0.675 to 0.717]) and FUNC score (c-index 0.701 [95% CI- 0.698 to 0.723], 0.668 [95% CI -0.657 to 0.680] and 0.727 [95% CI -0.708 to 0.753]). In the external independent Yale-ICH cohort, similar performance metrics were obtained for disability and severe disability (c-index 0.725 [95% CI -0.673 to 0.781] and 0.747 [95% CI -0.676 to 0.807], respectively). Similar AUC of predicting each outcome at 6 months, 1 and 2 years after ICH was achieved compared with ICH score and FUNC score. We developed a generalizable deep learning model to predict onset of dependent living and disability after ICH, which could help to guide treatment decisions, advise relatives in the acute setting, optimize rehabilitation strategies, and anticipate long-term care needs.

An Explainable and Generalizable Recurrent Neural Network Approach for Differentiating Human Brain States on EEG Dataset.

Electroencephalogram (EEG) is one of the most widely used brain computer interface (BCI) approaches. Despite the success of existing EEG approaches in brain state recognition studies, it is still challenging to differentiate brain states via explainable and generalizable deep learning approaches. In other words, how to explore meaningful and distinguishing features and how to overcome the huge variability and overfitting problem still need to be further studied. To alleviate these challenges, in this work, a multiple random fragment search-based multilayer recurrent neural network (MRFS-MRNN) is proposed to improve the differentiating performance and explore meaningful patterns. Specifically, an explainable MRNN module is proposed to capture the temporal dependences preserved in EEG time series. Besides, a MRFS module is designed to cut multiple random fragments from the entire EEG signal time course to improve the effectiveness of brain state differentiating ability. MRFS-MRNN is concatenatedto effectively overcome the huge variabilities and overfitting problems. Experiment results demonstrate that the proposed MRFS-MRNN model not only has excellent differentiating performance, but also has good explanation and generalization ability. The classification accuracies reach as high as 95.18% for binary classification and 89.19% for four-category classification on the individual level. Similarly, 95.53% and 85.84% classification accuracies are obtained for the binary and four-category classification on the group level. What's more, 94.28% and 85.43% classification accuracies of binary and four-category classifications are achieved for predicting brand new subjects. The experiment results showed that the proposed method outperformed other state-of-the-art (SOTA) models on the same underlying data and improved the explanation and generalization ability.

Editorial Special Issue on Explainable and Generalizable Deep Learning for Medical Imaging

Editorial Special Issue on Explainable and Generalizable Deep Learning for Medical Imaging

Assessing the potential of BirdNET to infer European bird communities from large-scale ecoacoustic data

1.Passive acoustic monitoring has become increasingly popular as a practical and cost-effective way of obtaining highly reliable acoustic data in ecological research projects. Increased ease of collecting these data means that, currently, the main bottleneck in ecoacoustic monitoring projects is often the time required for the manual analysis of passively collected recordings. In this study we evaluate the potential and current limitations of BirdNET-Analyzer v2.4, the most advanced and generic deep learning algorithm for bird recognition to date, as a tool to assess bird community composition through the automated analysis of large-scale ecoacoustic data.2.To this end, we study 3 acoustic datasets comprising a total of 629 environmental soundscapes collected in 194 different sites spread across a 19° latitude span in Europe. We analyze these soundscapes using both BirdNET and manual listening by local expert birders, and we then compare the results obtained through the two methods to evaluate the performance of the algorithm both at the level of each single vocalization and for entire recording sequences (1, 5 or 10 min).3.Since BirdNET provides a confidence score for each identification, minimum confidence thresholds can be used to filter out identifications with low scores, thus retaining only the most reliable ones. The volume of ecoacoustic data used in this study did not allow us to estimate species-specific minimum confidence thresholds for most taxa, so we instead used and evaluated global confidence thresholds selected for optimized results when consistently applied across all species.4.Our analyses reveal that BirdNET identifications can be highly reliable if a sufficiently high minimum confidence threshold is used. However, the inevitable trade-off between precision and recall does not allow to obtain satisfactory results for both metrics at the same time. We found that F1-scores remain moderate (<0.5) for all datasets and confidence thresholds studied, and that acoustic datasets of extended duration seem to be currently necessary for BirdNET to provide a reliable and minimally comprehensive picture of the target bird community. We estimate, however, that the usage of species- and context-specific minimum confidence thresholds would allow to substantially improve the global performance benchmarks obtained in this study.5.We conclude that a judicious use of AI-based identifications provided by BirdNET can represent a powerful method to assist in the assessment of bird community composition through the automated analysis of ecoacoustic data, especially when applied to acoustic datasets of extended duration.

Brain–machine interface based on deep learning to control asynchronously a lower-limb robotic exoskeleton: a case-of-study

BackgroundThis research focused on the development of a motor imagery (MI) based brain–machine interface (BMI) using deep learning algorithms to control a lower-limb robotic exoskeleton. The study aimed to overcome the limitations of traditional BMI approaches by leveraging the advantages of deep learning, such as automated feature extraction and transfer learning. The experimental protocol to evaluate the BMI was designed as asynchronous, allowing subjects to perform mental tasks at their own will.MethodsA total of five healthy able-bodied subjects were enrolled in this study to participate in a series of experimental sessions. The brain signals from two of these sessions were used to develop a generic deep learning model through transfer learning. Subsequently, this model was fine-tuned during the remaining sessions and subjected to evaluation. Three distinct deep learning approaches were compared: one that did not undergo fine-tuning, another that fine-tuned all layers of the model, and a third one that fine-tuned only the last three layers. The evaluation phase involved the exclusive closed-loop control of the exoskeleton device by the participants’ neural activity using the second deep learning approach for the decoding.ResultsThe three deep learning approaches were assessed in comparison to an approach based on spatial features that was trained for each subject and experimental session, demonstrating their superior performance. Interestingly, the deep learning approach without fine-tuning achieved comparable performance to the features-based approach, indicating that a generic model trained on data from different individuals and previous sessions can yield similar efficacy. Among the three deep learning approaches compared, fine-tuning all layer weights demonstrated the highest performance.ConclusionThis research represents an initial stride toward future calibration-free methods. Despite the efforts to diminish calibration time by leveraging data from other subjects, complete elimination proved unattainable. The study’s discoveries hold notable significance for advancing calibration-free approaches, offering the promise of minimizing the need for training trials. Furthermore, the experimental evaluation protocol employed in this study aimed to replicate real-life scenarios, granting participants a higher degree of autonomy in decision-making regarding actions such as walking or stopping gait.

Enhancing smart road safety with federated learning for Near Crash Detection to advance the development of the Internet of Vehicles

We introduce an innovative methodology for the identification of vehicular collisions within Internet of Vehicles (IoV) applications. This approach combines a knowledge base system with deep learning for model selection in an ensemble learning setting. It is designed to provide a general near-crash detection capability without relying on domain-specific knowledge, enabling the development of generic deep learning models. Our proposed methodology employs a novel deep learning approach, wherein multiple learning models are individually trained for each image. Subsequently, visual features are computed and stored for each trained image, along with the associated loss values from the training phase. This stored information is utilized to select the most suitable models for processing new image data during the testing phase. To facilitate efficient model selection, we employ a kNN (k Nearest Neighbors) strategy. To enhance both data and model security in IoV environments, we implement an intelligent federated learning (FL) strategy. Users are organized into clusters, and we employ two distinct aggregation methods, departing from conventional federated learning approaches. In the initial stage, we aggregate model data from all users to create a global model representing collective knowledge. In the subsequent stage, we aggregate models from each cluster to generate customized local models. Users are provided with both global and local models, allowing them to select the most suitable model for their specific crash detection needs. We test our approach, that we call Knowledge Guided Deep Learning for Near Crash Detection (KGDL-NCD), on well-known NCD benchmarks. The results demonstrate that KGDL-NCD surpasses baseline solutions, achieving an AUC (Area Under Curve) metric of 0.95.