ResNet Model Research Articles

Enhanced Vision Transformer with Custom Attention Mechanism for Automated Idiopathic Scoliosis Classification.

Scoliosis is a three-dimensional spinal deformity that is the most common among spinal deformities and causes extremely serious posture disorders in advanced stages. Scoliosis can lead to various health problems, including pain, respiratory dysfunction, heart problems, mental health disorders, stress, and emotional difficulties. The current gold standard for grading scoliosis and planning treatment is based on the Cobb angle measurement on X-rays. The Cobb angle measurement is performed by physical medicine and rehabilitation specialists, orthopedists, radiologists, etc., in branches dealing with the musculoskeletal system. Manual calculation of the Cobb angle for this process is subjective and takes more time. Deep learning-based systems that can evaluate the Cobb angle objectively have been frequently used recently. In this article, we propose an enhanced ViT that allows doctors to evaluate the diagnosis of scoliosis more objectively without wasting time. The proposed model uses a custom attention mechanism instead of the standard multi-head attention mechanism for the ViT model. A dataset with 7 different classes was obtained from 1456 patients in total from Elazığ Fethi Sekin City Hospital Physical Medicine and Rehabilitation Clinic. Multiple models were used to compare the proposed architecture in the classification of scoliosis disease. The proposed improved ViT architecture exhibited the best performance with 95.21% accuracy. This result shows that a superior classification success was achieved compared to ResNet50, Swin Transformer, and standard ViT models.

A feasibility Study of Detecting Chironomidae Larva in Water Treatment Filtration Processes using ResNet-based Image Recognition Deep Learning

There is a possibility that Chironomidae Larva may appear in sand and activated carbon filters during drinking water treatment. This study was conducted to determine whether the presence or absence of larva that may appear in filters through image data analysis. Image data were created for cases with and without larva background with interference materials such as sand and activated carbon granules used in the actual water treatment process. We used ResNet, one of the image classification deep learning models, and verified and evaluated its accuracy. Among the 12 models, the top three models with high TPR were ResNet50 No-pretrained LR=0.1, ResNet18 No-pretrained LR=0.001, and ResNet18 No-pretrained LR=0.01. Models trained by only obtaining the structure of the ResNet model without pre-training showed higher accuracy and superior performance. Among the models with the highest accuracy, ResNet50 No-pretrained LR=0.1 had a large TPR value. However, FPR also showed a large value. Therefore, it could not be suitable for judging the presence or absence of larva in the water treatment process.When comparing ResNet18 No-pretrained LR=0.001 and ResNet18 No-pretrained LR=0.01, which had the second and third highest TPRs, the ResNet18 No-pretrained LR=0.01 model had the highest Accuracy and F1 Score.

ResT-IMU: A Two-Stage ResNet-Transformer Framework for Inertial Measurement Unit Localization

To address the challenges of accurate indoor positioning in complex environments, this paper proposes a two-stage indoor positioning method, ResT-IMU, which integrates the ResNet and Transformer architectures. The method initially processes the IMU data using Kalman filtering, followed by the application of windowing to the data. Residual networks are then employed to extract motion features by learning the residual mapping of the input data, which enhances the model’s ability to capture motion changes and predict instantaneous velocity. Subsequently, the self-attention mechanism of the Transformer is utilized to capture the temporal features of the IMU data, thereby refining the estimation of movement direction in conjunction with the velocity predictions. Finally, a fully connected layer outputs the predicted velocity and direction, which are used to calculate the trajectory. During training, the RMSE loss is used to optimize velocity prediction, while the cosine similarity loss is employed for direction prediction. Theexperimental results demonstrate that ResT-IMU achieves velocity prediction errors of 0.0182 m/s on the iIMU-TD dataset and 0.014 m/s on the RoNIN dataset. Compared with the ResNet model, ResT-IMU achieves reductions of 0.19 m in ATE and 0.05 m in RTE on the RoNIN dataset. Compared with the IMUNet model, ResT-IMU achieves reductions of 0.61 m in ATE and 0.02 m in RTE on the iIMU-TD dataset and reductions of 0.32 m in ATE and 0.33 m in RTE on the RoNIN dataset. Compared with the ResMixer model, ResT-IMU achieves reductions of 0.13 m in ATE and 0.02 m in RTE on the RoNIN dataset. These improvements indicate that ResT-IMU offers superior accuracy and robustness in trajectory prediction.

Optimising CNN Architecture for Accurate Detection of Tessellated Retinal Disease Using Fundus Images

Eyes are one of the vital organs for human beings, which serve as a primary gateway to perceive the surroundings. An abnormal situation, namely tessellated eye, is commonly caused by myopia, which has a characteristic mosaic-like pattern that can cause early vision loss, particularly in infants and youngsters. This work contributes with the usage of a variety of deep learning models to diagnose tessellated and normal fundus images automatically which will improve early detection that can lead to the prevention of vision loss. This study uses a standard dataset of 732 annotated fundus images obtained from Mendeley, Kaggle and a local ophthalmology centre. It also uses a variety of Convolutional Neural Network (CNN) architectures, including VGG16, VGG19, ResNet50 and sequential models, that are experimented for determining the best model and examined. Initially, the fundus images are pre-processed and enhanced to improve model resilience. Out of all the architectures, ResNet50 outperformed as the best model, with an accuracy of 79.45%, while VGG16 with data augmentation reported the best accuracy of 90.8%. Grad-CAM (Gradient-weighted Class Activation Mapping), an Explainable Artificial Intelligence (XAI) mechanism, is used to create heatmaps for interpretability, emphasising spots and pathologies that contribute to the model’s experimentation and judgements. The outcomes of this research highlight potential models namely ResNet50 and augmented VGG16 for reliably diagnosing the fundus images as tessellated or normal. The study also seeks to serve as a platform for future investigation of classifying various automated retinal diseases.

Deep learning-based classification of speech disorder in stroke and hearing impairment.

Speech disorders can arise from various causes, including congenital conditions, neurological damage, diseases, and other disorders. Traditionally, medical professionals have used changes in voice to diagnose the underlying causes of these disorders. With the advancement of artificial intelligence (AI), new possibilities have emerged in this field. However, most existing studies primarily focus on comparing voice data between normal individuals and those with speech disorders. Research that classifies the causes of these disorders within the abnormal voice data, attributing them to specific etiologies, remains limited. Therefore, our objective was to classify the specific causes of speech disorders from voice data resulting from various conditions, such as stroke and hearing impairments (HI). We experimentally developed a deep learning model to analyze Korean speech disorder voice data caused by stroke and HI. Our goal was to classify the disorders caused by these specific conditions. To achieve effective classification, we employed the ResNet-18, Inception V3, and SEResNeXt-18 models for feature extraction and training processes. The models demonstrated promising results, with area under the curve (AUC) values of 0.839 for ResNet-18, 0.913 for Inception V3, and 0.906 for SEResNeXt-18, respectively. These outcomes suggest the feasibility of using AI to efficiently classify the origins of speech disorders through the analysis of voice data.

Classificação de Aves Predadoras: Fine-tuning Progressivo em Redes Neurais Convolucionais

AbstractThis paper presents a process for classifying predatory birds byfamily and species. The motivation for this study arises from thehigh variability observed among birds of different species and theimportance of performing classification efficiently and in a timelymanner. Additionally, this work aims to analyze the impact of usingRGB channels in comparison to grayscale images on classificationperformance, as well as the effect of applying data augmentationtechniques during training. The dataset contains 42,475 images,distributed across 6 families and 41 species. The process employsfine-tuning, using the ResNet-50 model. Early stopping was appliedto control overfitting and obtain the best model. The test resultshighlight the effectiveness of the proposed process in classificationtasks, with performance varying across different input configurations.For species classification, the model trained with grayscalechannels achieved an F1-Score of 0.80. Using RGB channels improvedthe performance significantly, resulting in an F1-Score of0,86. Further applying data augmentation techniques to the RGBslightly improved the metrics, achieving an F1-Score of 0.87. Theseresults demonstrate the benefits of incorporating color informationand data augmentation in enhancing classification accuracy.

Enhanced Deep Residual Network based Self-Learning framework for Mango leaf disease Classification: Focus on Anthracnose and Grey Blight

In modern agriculture, ensuring early and accurate diagnosis of plant diseases is vital for crop health and yield. Deep learning has increasingly been utilized for diagnosing diseases in mango leaves using pathological images. Most existing solutions rely heavily on supervised learning models, which require extensive labeled data—a process that is both time-consuming and labor-intensive for agricultural experts. To ease this burden, a self-supervised learning approach has been developed that depends on minimal labeled data. This proposed work introduces a semi-supervised learning model for classifying mango leaf diseases, reducing the need for exhaustive manual annotation. The system is trained using 3,654 images of diseased mango leaves. BYOL (Bootstrap Your Own Latent), a self-supervised algorithm, was employed to train a ResNet with SE blocks network, enabling it to extract meaningful features from infected regions without relying entirely on labeled data. With only 30% of the dataset labeled, a self-supervised learning approach was used to develop a classification model for identifying Anthracnose, Grey Blight, and healthy leaves. This technique achieved an impressive classification accuracy of 98.11%, slightly surpassing the fully supervised ResNet-50 model's accuracy of 97.62%. The outcome demonstrates that accurate disease detection in mango leaves can be accomplished with reduced labeling effort, supporting more efficient and scalable agricultural diagnostics.

Optimal deep learning based vehicle detection and classification using chaotic equilibrium optimization algorithm in remote sensing imagery

Remote sensing images (RSI), such as aerial or satellite images, produce a large-scale view of the Earth’s surface, which gets them used to track and monitor vehicles from several settings, like border control, disaster response, and urban traffic surveillance. Vehicle detection and classification using RSIs is a vital application of computer vision and image processing. It contains locating and identifying vehicles from the image. It is done using many approaches that have object detection approaches, namely YOLO, Faster R-CNN, or SSD, which utilize deep learning (DL) to locate and identify the image. Additionally, the classification of vehicles from RSIs contains classification of them based on their variety, such as trucks, motorcycles, cars or buses, utilizing machine learning (ML) techniques. This article designed and developed an automated vehicle type detection and classification using a chaotic equilibrium optimization algorithm with deep learning (VDTC-CEOADL) on high-resolution RSIs. The VDTC-CEOADL technique presented examines high-quality RSIs for the accurate detection and classification of vehicles. The VDTC-CEOADL technique employs a YOLO-HR object detector with a residual network as the backbone model to accomplish this. In addition, CEOA based hyperparameter optimizer is designed for the parameter tuning of the ResNet model. For the vehicle classification process, the VDTC-CEOADL technique exploits the attention-based long-short-term memory (ALSTM) mod-el. Performance validation of the VDTC-CEOADL technique is validated on a high-resolution RSI dataset, and the results portrayed the supremacy of the VDTC-CEOADL technique in terms of different measures.

Visual Rhythms based Video Face Spoofing Detection using Deformable Self Attention ResNet Model

Visual Rhythms based Video Face Spoofing Detection using Deformable Self Attention ResNet Model

Domain-Adaptive RFI Detection Using Fine-Tuned Time-Frequency Deep Models and Visual Explainability

Radio Frequency Interference (RFI) remains a significant threat to the reliability of modern wireless systems, particularly as signal environments grow increasingly diverse and congested. This paper introduces a novel domain-adaptive deep learning framework for robust RFI detection across heterogeneous wireless environments. Unlike existing approaches that use static pre-trained convolutional neural networks (CNNs), we propose a two-stage transfer learning strategy wherein ResNet50 and AlexNet models are selectively fine-tuned on domain-specific signal datasets represented as spectrograms and scalograms. These time-frequency transformations capture complementary spectral characteristics—spectrograms model persistent interference patterns, while scalograms highlight transient, bursty anomalies. The fine-tuned networks extract high-level semantic features that are then adaptively weighted using an attention mechanism, enabling the model to emphasize the most informative representations from each domain. The fused features are classified via a lightweight CNN, which balances accuracy with computational efficiency. To promote transparency and model trustworthiness, we further integrate Grad-CAM-based visual explanations that highlight the discriminative regions within the time-frequency maps responsible for the model’s decisions. Experimental evaluations across multiple signal domains, including synthetic and real-world datasets, demonstrate that the proposed approach not only achieves state-of-the-art accuracy (98.1%) but also generalizes effectively to unseen interference types. This framework offers a scalable, explainable, and transferable solution for real-time RFI detection in complex wireless, satellite, and edge-based IoT systems.

Enhancing Parking Systems with QR Code-Integrated Automatic License Plate Recognition through Convolutional Neural Networks

This abstract describes the development and evaluation of an Automatic License Plate Recognition (ALPR) system designed to simplify the process of parking ticket generation. The traditional paradigm of manual entry of license plate information by parking personnel for exiting vehicles is replaced by the automated system proposed in this study. The system integrates a YOLO (You Only Look Once) model for the automatic recognition of license plates in vehicle images. After this initial identification, a series of pre-processing and image segmentation techniques are applied to isolate and recognize the individual digits within the license plate. A ResNet model is then used to classify the license plates. The research focuses specifically on Malaysian license plates. The experimental results show that the YOLO model recognizes license plates robustly and accurately and has a high degree of reliability. However, when validating the data set, the ResNet model achieves an accuracy of around 80 %. The study points out inherent challenges, including potential errors in segmentation, problems with non-standardized or damaged tags, and the presence of digits that may have visual similarities. In summary, while the YOLO model is reliable in recognizing license plates, the classification accuracy of the ResNet model can be further improved. Overcoming challenges such as segmentation noise and variations in license plate conditions could further optimize the overall performance of the system.

Integrating ResNet-50 and Vision Transformer Architectures for Robust and Efficient Tomato Fruit Ripeness Classification

This article introduces a new hybrid deep learning framework that combines ResNet-50 and Vision Transformer (ViT) models to classify tomato fruits by ripeness and quality. The hybrid model takes advantage of the ResNet-50's ca-pability to extract features in local spatial regions. Additionally, it combines ResNet-50 with ViT's ability to establish a global contextual relationship be-tween elements with self-attention. The framework overcomes the limitations of utilizing a single model. The framework was trained and evaluated, using a balanced and diverse dataset. It consists of four tomato classes, ripe, unripe, unevenly ripened, and damaged, collected under controlled conditions in the field. The experiments conducted aggressively in the study demonstrated a high classification accuracy of 98% with a hybrid framework compared to individual ResNet-50 and ViT models. The results in addition to the general performance were validated further using precision, recall, the area under the curve (AUC) and the F1-score. The study also included a computational efficiency test to look for accuracy versus multiple time and spatial resource costs. The research suc-cessfully addresses the challenges with dataset diversity, computational costs, and real-time feature deployments through the conceptualization of strategies (data augmentation, transfer learning) to improve generalization. The hybrid framework is expected to provide similar design principles and measures for use with other agricultural products. It can help in real-time sorting applications in variable real-world scenarios in agricultural operations. The research pro-vides a pathway for developing smarter, scalable, and sustainable solutions in food quality and safety systems within precision agriculture. Future work will incorporate deployment and optimize the framework for edge devices in a re-al-world variable farm environment.

Small sample cross-domain bearing fault diagnosis method based on signal denoising lightweight model

Abstract During the operation of bearings, their structural parameters, loads, signal transmission paths, etc, are highly dynamic, resulting in obvious distribution differences in the statistical characteristics of the data. In addition, the real fault data set is often too small and lacks fault labels, resulting in poor performance of data-driven neural networks in small sample cross-domain fault diagnosis. In order to solve the above problems faced by engineering practice and achieve high-precision fault diagnosis, a signal denoising lightweight (FMECR-18) model for small sample cross-domain bearing fault diagnosis is proposed. First, the two-domain signal data is denoised using a denoising module (FME) based on the combination of eigenmode decomposition (FMD) and multiscale entropy screening (MSE), and the denoised vibration signal is upgraded to a multi-dimensional signal through channel expansion. Next, the basic structure of the ResNet18 model is changed by replacing the conventional convolution residual module with a deformable convolution residual module and a convolution attention module (CBAM) to create a CR-18 model. Using the fine-tuning strategy discussed in this paper, the ability of the FMECR-18 model to extract fault features for two types of data and the effectiveness of bearing fault classification are improved by training two types of data and adjusting the model settings. Using the fine-tuning strategy theory proposed in this paper, the fault feature extraction ability of the FMECR-18 model for dual-domain data and the bearing fault classification effect are improved through dual-domain data training and model freezing adjustment. Six cross-dataset migration tasks are constructed using 7 different public datasets. The average diagnostic results of the target domain test set of the FMECR-18 model are 99.89%, 99.70%, 99.61%, 98.18%, 96.91%, and 96.06%, respectively, indicating that the method proposed in this paper has achieved good results in cross-dataset migration tasks.

MRI-based 2.5D deep learning radiomics nomogram for the differentiation of benign versus malignant vertebral compression fractures.

Vertebral compression fractures (VCFs) represent a prevalent clinical problem, yet distinguishing acute benign variants from malignant pathological fractures constitutes a persistent diagnostic dilemma. To develop and validate a MRI-based nomogram combining clinical and deep learning radiomics (DLR) signatures for the differentiation of benign versus malignant vertebral compression fractures (VCFs). A retrospective cohort study was conducted involving 234 VCF patients, randomly allocated to training and testing sets at a 7:3 ratio. Radiomics (Rad) features were extracted using traditional Rad techniques, while 2.5-dimensional (2.5D) deep learning (DL) features were obtained using the ResNet50 model. These features were combined through feature fusion to construct deep learning radiomics (DLR) models. Through a feature fusion strategy, this study integrated eight machine learning architectures to construct a predictive framework, ultimately establishing a visualized risk assessment scale based on multimodal data (including clinical indicators and Rad features).The performance of the various models was evaluated using the receiver operating characteristic (ROC) curve. The standalone Rad model using ExtraTrees achieved AUC=0.801 (95%CI:0.693-0.909) in testing, while the DL model an AUC value of 0.805 (95% CI: 0.690-0.921) in the testing cohort. Compared with the Rad model and DL model, the performance superiority of the DLR model was demonstrated. Among all these models, the DLR model that employed ExtraTrees algorithm performed the best, with area under the curve (AUC) values of 0.971 (95% CI: 0.948-0.995) in thetraining dataset and 0.828 (95% CI: 0.727-0.929) in the testing dataset. The performance of this model was further improved when combined with clinical and MRI features to form the DLR nomogram (DLRN), achieving AUC values of 0.981 (95% CI: 0.964-0.998) in the training dataset and 0.871 (95% CI: 0.786-0.957) in the testing dataset. Our study integrates handcrafted radiomics, 2.5D deep learning features, and clinical data into a nomogram (DLRN). This approach not only enhances diagnostic accuracy but also provides superior clinical utility. The novel 2.5D DL framework and comprehensive feature fusion strategy represent significant advancements in the field, offering a robust tool for radiologists to differentiate benign from malignant VCFs.

AI-IoT based smart agriculture pivot for plant diseases detection and treatment

There are some key problems faced in modern agriculture that IoT-based smart farming. These problems such shortage of water, plant diseases, and pest attacks. Thus, artificial intelligence (AI) technology cooperates with the Internet of Things (IoT) toward developing the agriculture use cases and transforming the agriculture industry into robustness and ecologically conscious. Various IoT smart agriculture techniques are escalated in this field to solve these challenges such as drop irrigation, plant diseases detection, and pest detection. Several agriculture devices were installed to perform these techniques on the agriculture field such as drones and robotics but in expense of their limitations. This paper proposes an AI-IoT smart agriculture pivot as a good candidate for the plant diseases detection and treatment without the limitations of both drones and robotics. Thus, it presents a new IoT system architecture and a hardware pilot based on the existing central pivot to develop deep learning (DL) models for plant diseases detection across multiple crops and controlling their actuators for the plant diseases treatment. For the plant diseases detection, the paper augments a dataset of 25,940 images to classify 11-classes of plant leaves using a pre-trained ResNet50 model, which scores the testing accuracy of 99.8%, compared to other traditional works. Experimentally, the F1-score, Recall, and Precision, for ResNet50 model were 99.91%, 99.92%, and 100%, respectively.

Medicinal Plant Recognition

This paper introduces the concept of an “Medicinal Plant Recognition” has historically relied on expert knowledge and manual classification, often leading to errors and inefficiencies. This work presents an automated medicinal plant recognition system using deep learning techniques, specifically a fine-tuned ResNet50 model, coupled with a web- based user interface built with Flask. The system enhances classification accuracy by applying a series of preprocessing steps including grayscale conversion, edge detection, thresholding, and sharpening. Users can interact with the application by uploading leaf images or using real-time camera input for immediate predictions. Upon classification, the system provides the plant's name, the confidence score, and detailed medicinal information such as benefits and potential side effects. Designed for scalability, the platform bridges traditional medicinal knowledge with modern artificial intelligence approaches, offering applications in healthcare, agriculture, education, and biodiversity conservation.

Artificial Intelligence Model Assists Knee Osteoarthritis Diagnosis via Determination of K-L Grade.

Background: Knee osteoarthritis (KOA) affects 37% of individuals aged ≥ 60 years in the national health survey, causing pain, discomfort, and reduced functional independence. Methods: This study aims to automate the assessment of KOA severity by training deep learning models using the Kellgren-Lawrence grading system (class 0~4). A total of 15,000 images were used, with 3000 images collected for each grade. The learning models utilized were DenseNet201, ResNet101, and EfficientNetV2, and their performance in lesion classification was evaluated and compared. Statistical metrics, including accuracy, precision, recall, and F1-score, were employed to assess the feasibility of applying deep learning models for KOA classification. Results: Among these four metrics, DenseNet201 achieved the highest performance, while the ResNet101 model recorded the lowest. DenseNet201 demonstrated the best performance with an overall accuracy of 73%. The model's accuracy by K-L grade was 80.7% for K-L Grade 0, 53.7% for K-L Grade 1, 72.7% for K-L Grade 2, 75.3% for K-L Grade 3, and 82.7% for K-L Grade 4. The model achieved a precision of 73.2%, a recall of 73%, and an F1-score of 72.7%. Conclusions: These results highlight the potential of deep learning models for assisting specialists in diagnosing the severity of KOA by automatically assigning K-L grades to patient data.

Enhancing segmentation accuracy of the common iliac vein in OLIF51 surgery in intraoperative endoscopic video through gamma correction: a deep learning approach.

The principal objective of this study was to develop and evaluate a deep learning model for segmenting the common iliac vein (CIV) from intraoperative endoscopic videos during oblique lateral interbody fusion for L5/S1 (OLIF51), a minimally invasive surgical procedure for degenerative lumbosacral spine diseases. The study aimed to address the challenge of intraoperative differentiation of the CIV from surrounding tissues to minimize the risk of vascular damage during the surgery. We employed two convolutional neural network (CNN) architectures: U-Net and U-Net++ with a ResNet18 backbone, for semantic segmentation. Gamma correction was applied during image preprocessing to improve luminance contrast between the CIV and adjacent tissues. We used a dataset of 614 endoscopic images from OLIF51 surgeries for model training, validation, and testing. The U-Net++/ResNet18 model outperformed, achieving a Dice score of 0.70, indicating superior ability in delineating the position and shape of the CIV compared to the U-Net/ResNet18 model, which achieved a Dice score of 0.59. Gamma correction increased the differentiation between the CIV and the artery, improving the Dice score from 0.44 to 0.70. The findings demonstrate that deep learning models, especially the U-Net++ with ResNet18 enhanced by gamma correction preprocessing, can effectively segment the CIV in intraoperative videos. This approach has the potential to significantly improve intraoperative assistance and reduce the risk of vascular injury during OLIF51 procedures, despite the need for further research and refinement of the model for clinical application.

ATR-FTIR combined with chemometrics to distinguish geographical indications from non-geographical indications Gastrodia elata Bl

ATR-FTIR combined with chemometrics to distinguish geographical indications from non-geographical indications <i>Gastrodia elata</i> Bl

Deep Learning-Enhanced Spectrogram Analysis for Anatomical Region Classification in Biomedical Signals

Accurate classification of biomedical signals is essential for advancing non-invasive diagnostic techniques and improving clinical decision-making. This study introduces a deep learning-augmented spectrogram analysis framework for classifying biomedical signals into eight anatomically distinct regions, thereby addressing a significant deficiency in automated signal interpretation. The proposed approach leverages a fine-tuned ResNet50 model, pre-trained on ImageNet, and adapted for a single-channel spectrogram input to ensure robust feature extraction and high classification accuracy. Spectrograms derived from palpation and percussion signals were preprocessed into grayscale images and optimized through data augmentation and hyperparameter tuning to enhance the model’s generalization. The experimental results demonstrate a classification accuracy of 93.37%, surpassing that of conventional methods and highlighting the effectiveness of deep learning in biomedical signal processing. This study bridges the gap between machine learning and clinical applications, enabling an interpretable and region-specific classification system that enhances diagnostic precision. Future work will explore cross-domain generalization, multi-modal medical data integration, and real-time deployment for clinical applications. The findings establish a significant advancement in non-invasive diagnostics, demonstrating the potential of deep learning to refine and automate biomedical signal analysis in clinical practice.