Background: Enchondroma of the metacarpal can compromise bone integrity and hand function, requiring surgical excision and reconstruction. This study presents a patient-specific computational-to-clinical workflow for the design, fabrication, and clinical evaluation of a biodegradable poly-L-lactic acid (PLLA) scaffold for reconstruction of a fifth metacarpal defect following enchondroma excision. Methods: Bilateral computed tomography (CT) scans were acquired, and the healthy contralateral metacarpal was mirrored to generate a patient-specific anatomical model. A hollow modular PLLA scaffold was designed using computer-aided design, mechanically optimized through ASTM D695 compression testing and finite element analysis (FEA), and fabricated using fused deposition modeling. Following tumor excision, the scaffold was implanted with autologous bone marrow aspirate and osteoconductive graft material and stabilized using a K-wire. Clinical follow-up included serial radiographic assessment and scanning electron microscopy (SEM) analysis of a retrieved scaffold specimen after 18 months. Results: Mechanical testing and FEA identified a 2[Formula: see text]mm wall thickness as the optimum scaffold configuration, with numerical predictions agreeing with experimental results within [Formula: see text]. Radiographic follow-up at 31 and 46 weeks demonstrated progressive mineralization and osteointegration within the scaffold cavity. SEM analysis of the scaffold retrieved after 18 months revealed surface erosion and a residual wall thickness of approximately 733[Formula: see text][Formula: see text]m, corresponding to an approximately 63% reduction from the original 2[Formula: see text]mm wall thickness, confirming controlled in vivo biodegradation while maintaining structural integrity. Conclusions: The proposed patient-specific PLLA scaffold and computational-to-clinical workflow demonstrated preliminary feasibility for reconstruction of metacarpal bone defects following benign tumor excision. The scaffold provided temporary mechanical support while facilitating guided bone regeneration and progressive material resorption. This integrated design and evaluation strategy may provide a promising framework for the development of personalized biodegradable orthopedic implants and warrants further validation through larger clinical studies.
Laryngeal cancer is a condition where the cancer cells form in the larynx tissues. Laryngeal cancer is caused by the use of tobacco and alcohol consumption, which are known to increase the risk of developing laryngeal cancer. The symptoms of laryngeal cancer are a persistent sore throat and ear pain. Several limitations are faced by previous techniques, including high diagnostic inconsistency, speech signal variability, limited real-time diagnosis, and the cost of technology. To overcome these limitations, a model named Deep Residual Quantum Network (Deep RQN) is developed for classifying laryngeal cancer. Initially, the input speech signal is subjected to signal preprocessing, which is performed using an adaptive Gaussian filter. Then, feature extraction is accomplished for extracting the features that include Zero Crossing Rate (ZCR), glottal waveform, Discrete Wavelet Transform (DWT), Mel Frequency Ceptral Coefficient (MFCC), and statistical features, including standard deviation, skewness, and kurtosis. Then, the extracted features are subjected to a Vocal Tract Support system to classify laryngeal cancer, which is done by Deep RQN, where the result is obtained as normal or disordered. When the result is considered normal, the corresponding signal is saved into the database, and if it is assumed to be the disordered signal, then it undergoes signal segmentation and speech enhancement. The signal segmentation is performed using Maximum A Posteriori Probability (MAP). Moreover, Natural Language Processing (NLP) based speech enhancement is done using the Gaussian Mixture Model (GMM), and its output results in a word sequence. Furthermore, the enhanced signal is applied to the database. The Deep RQN combines a Deep Quantum Neural Network (DQNN) and a Deep Residual Network (DRN). The Deep RQN attained superior outcomes than previous methods with metrics, including accuracy, sensitivity, and specificity of 91.55%, 91.90%, and 92.56%, respectively.
Spinal Cord Injury (SCI) detection is crucial in medical imaging for early analysis and effective therapy planning. This paper emphasizes the role of deep learning (DL) in enhancing diagnostic efficiency. This paper introduces an innovative hybrid DL model for classifying SCI levels. The proposed model, called the Shepherd Siamese Convolutional Network (ShSCN-Net), combines the Siamese Convolutional Neural Network (SCNN) and the Shepherd Convolutional Neural Network (ShCNN). First, the Computed Tomography (CT) image obtained from a database is given as the input to the image denoising phase, where the denoising of the images is done by using a median filter. Then the denoised images are segmented using a Multi-scale Attention Network (MANet). Later, using the active contour model, the disc localization of the spinal cord image is done. Subsequently, feature extraction is done to cut down the magnitude of the input information for processing. Next, injury level detection is achieved through the proposed model, with the identified injuries classified into four categories: Cervical (C1-C8), Thoracic (T1-T12), Lumbar (L1-L5), and Sacral (S1-S5) through the ShSCN-Net. The proposed ShSCN-Net is evaluated for its effectiveness in determining spinal cord injuries using various metrics, including accuracy, True Positive Rate (TPR), and True Negative Rate (TNR). Across 90% of the data used for training, the ShSCN-Net achieves notable values of 91.042% for accuracy, 92.855% for TPR, and 91.644% for TNR.
In the current advanced universe, Heart Disease (HD) is considered the most dangerous. Since this disease affects a person very quickly, they have little time to receive treatment. Therefore, accurately and promptly examining patients is a major challenge for medical institutions. A poor examination by the hospital can lead to negative opinions and damage its reputation. At the same time, the cost of treatment becomes high, making it unaffordable for many patients. To improve heart disease detection, a hybrid deep learning model, Xception Fused Residual SqueezeNet (XcepFRSNet), is developed. From the chosen database, the input data is standardized through z-score normalization, and missing entries are imputed to prepare it for subsequent processing. The Wave-Hedges metric is employed for feature selection, and bootstrapping is subsequently applied to augment the data. The proposed XcepFRSNet model performs heart disease prediction, incorporating layer modifications guided by the Taylor concept. The proposed XcepFRSNet demonstrates strong results, attaining an accuracy of 90.942%, a sensitivity of 91.345%, and specificity of 89.67%.
Alzheimer’s Disease (AD) is a neurological condition that affects a large number of people.”Brain atrophy is a result of this neurological condition, which can result in memory loss, cognitive disabilities, and death. In its early stages, AD is difficult to identify. Thus, an early diagnosis and effective treatment of AD are more beneficial and create fewer complexities. AD is a common brain condition that is hard to recognize, and its classification procedures require a biased representation of traits to distinguish similar brain patterns. Multimodal neuro-image integrates numerous clinical images, which assist in identifying and diagnosing AD with better accuracy and effectiveness. Here, an efficient deep learning-based technique is developed to detect AD from“the multi-modal data.”Online resources are used to collect the necessary multimodal data, which includes Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET), and genetic data in the beginning. In this study, the developed model used 89 affected subjects and 90 healthy control subjects. Then, the collected MRI, PET, and genetic data are fed into the preprocessing stage. Here, genetic data is pre-processed using the data-filling approach. Likewise, the “MRI and PET images are used for the Regions of Interest (ROI) segmentation phase, and it is executed via Dilated TransUNet (DTUNet). To reduce processing time, the ROI segmentation process removes specific regions from MRI and PET images. The pre-processed MRI, PET, and genetic data are passed to the Graph Convolutional Networks (GCNs) to extract the features. From the extracted features, the relevant features are selected using the Enhanced Secretary Bird Optimization (ESBO) algorithm, and this feature is multiplied by the optimal weight selected that is selected using the same ESBO to form the weighted fused features. The final stage involves giving the weighted fused features to the Adaptive Deep CapsNet (ADCapsNet) to diagnose AD. Here, the effectiveness of the ADCapsNet is also enhanced by tuning the parameters using ESBO. Finally, the analysis procedures are executed to prove the proposed framework’s efficiency. In the evaluation, the accuracy of the developed model is 95.17, which is 7.11%, 2.3%, 4.3%, and 1.46% better than “Convolutional Neural Networks (CNNs), Residual Neural Network (ResNet), Visual Geometry Group 16 (VGG-16), and Deep CapsNet (DcapsNet).” The developed approach’s ability to detect AD at an early stage is demonstrated in the experimental outcome, which allows for better planning and improved quality of life for individuals.
Cardiac Arrhythmia, which encompasses irregular heart rhythms such as tachycardia, bradycardia, ectopic pulses, represents a significant cause of cardiovascular morbidity and mortality worldwide, caused by disruptions in the heart's electrical conduction system. Accurate and automated interpretation of the Electrocardiogram (ECG) is consequently crucial for the timely diagnosis and selection of treatment. This study introduces CadCNN 1.0, a novel hybrid Deep Learning (DL) framework that combines one-dimensional Convolutional Neural Networks (1D-CNN) for spatial feature extraction with Bidirectional Long Short-Term Memory (BiL-STM) for temporal sequence modeling. This methodology enables the accurate characterization of ECG waveforms and the model was trained and validated using the MIT-BIH Arrhythmia Database, classifying ECG signals into five categories: Normal (N), Fusion (F), Supraventricular Ectopic Beat (SVEB), Ventricular Ectopic Beat (VEB), and Unknown (Q) Beat. The Gray Wolf Optimization (GWO) algorithm was employed to optimize feature selection and model convergence, resulting in exceptional diagnostic performance with an Accuracy of 99.92%, Sensitivity (Recall) of 99.81%, Specificity of 99.93%, Precision of 99.92%, F1-score of 99.92% along with MCC of 99.84% with Number of Wolves (Nw) of 10 represents the number of wolves in the population, corresponding to the number of candidate feature subsets simultaneously explored in each iteration, that enhances population diversity and prevents premature convergence and Max Iteration of 10 which denotes the maximum number of optimization cycles during which the wolves update their positions toward the optimal feature subset, ensuring iterative refinement, stability of global optimum. The confusion matrix demonstrates CadCNN 1.0's discriminative ability to differentiate morphologically equivalent Arrhythmic classes with near-perfect precision, as evidenced by its strong diagonal dominance and minimal off-diagonal misclassifications. Furthermore, the Breast Cancer Dataset performed cross-domain validation, resulting in an Accuracy of 97.50%, Sensitivity (Recall) of 96.49%, Specificity of 97.18%, Precision of 96.49%, F1-score of 96.49%, MCC of 92.53%. These results underscore the model's adaptability to heterogeneous biomedical data. CadCNN 1.0 provides a clinically reliable, generalizable, computationally efficient computer-aided diagnostic system that advances automated Arrhythmia detection and enables intelligent, patient-centered healthcare diagnostics by integrating spatial-temporal feature learning, meta-heuristic optimization, and interpretable confusion-matrix validation.
The Cup-to-Disc Ratio (CDR) is a crucial diagnostic measure for glaucoma, comparing the diameters of the Optic Disc (OD) and Optic Cup (OC) in retinal images. Manual CDR measurement can be inaccurate due to the irregular shapes of these structures, leading to potential diagnostic errors. Automated semantic segmentation offers a faster, more accurate alternative for CDR assessment. This study aims to develop a novel architecture, ARS-Former (Attention-Refinement SegFormer), to improve the semantic segmentation of retinal images. This study focuses on improving the segmentation of small, complex structures, such as the OD and OC. ARS-Former is a modified SegFormer that integrates a Convolutional Block Attention Module (CBAM) into the encoder skip connections and adds a Refinement Block-ASPP in the decoder. The CBAM integration enhances the representation of essential features, channel-wise and spatially, ensuring that the decoder receives more relevant information. The Refinement Block-ASPP improves the model’s ability to capture multiscale context and refine object boundaries, thereby improving segmentation. Experimental results on the Drishti_GS dataset demonstrate that ARS-Former achieves 99% accuracy, 98% precision, 97.3% recall, 95.4% IoU, 97.6% DSC, and 98% ROC. These results indicate the architecture’s excellent performance and robustness in segmenting the OD and OC. The proposed method demonstrates excellent performance and robustness in OD and OC segmentation. However, further improvements are required, particularly in handling lower-quality images, to enhance robustness across diverse imaging conditions and ensure applicability in actual clinical practice.
The goal of this research is to create a reliable and effective deep learning framework for automatically identifying brain cancers from magnetic resonance imaging (MRI) scans. Brain tumors typically result from the unchecked proliferation of brain cells, some of which may develop into cancer. The ability of radiologists to accurately identify, segment, and classify tumor areas in MRI images is crucial, as it can result in inconsistent diagnoses. This study proposes a novel Transfer Learning-based Normalized Mixed Pooling SqueezeNet for Brain Tumor Detection (TLNMPS-BTD) architecture to improve tumor classification accuracy and reliability. A publicly accessible BraTS brain MRI dataset was used to assess the suggested framework, guaranteeing the repeatability and dependability of the experimental findings. In order to increase picture quality for further analysis, the suggested method first preprocesses MRI brain images using the Wiener Filtering Technique to eliminate noise and enhance image contrast. The region of interest is subsequently extracted by efficient segmentation using the Attention Mechanism-assisted SegNet (AMS). New Texture Computation on Local Gabor Transitional Pattern is a revolutionary texture-based technique for feature extraction. It increases the discriminative capability of retrieved features and improves low-resolution MRI pictures. The lightweight Normalized Mixed Pooling SqueezeNet (TLNMPS) model, which is based on Transfer Learning, receives these extracted features for tumor identification and classification. The proposed TLNMPS-BTD framework outperformed current state-of-the-art methods, achieving an accuracy of 96.8% and an F1-score of 96.7% at 90% training data.
A Deep Learning (DL)-based framework for Diabetic Retinopathy (DR) classification, named the Student Feedback Artificial Tree Algorithm-based Recurrent Fuzzy Long Short-Term Memory Network (SFATA-based RFLSTMN), is proposed in this study. First, the input image undergoes pre-processing using a Gaussian filter combined with Contrast-Limited Adaptive Histogram Equalization (CLAHE). Next, lesion segmentation is performed through deep joint+U-Net, taking intensity-based distance into account. Subsequently, the feature extraction stage identifies and extracts the relevant features. Finally, DR classification is conducted using the RFLSTMN model, which integrates the strengths of Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) architectures. Furthermore, RFLSTMN is trained by the Student Feedback Artificial Tree Algorithm (SFATA), which is developed through the integration of the Feedback Artificial Tree Algorithm (FATA) and Student Psychology-Based Optimization (SPBO). The performance assessment shows that the SFATA-based RFLSTMN model achieved superior results across all evaluation metrics, attaining an accuracy of 92.3%, sensitivity of 92.9%, specificity of 92.1%, precision of 91.5%, and F1-score of 92.2%.
Glaucoma, a major cause of irreversible blindness, requires accurate and early detection of optic disc (OD) and optic cup (OC) deformations in fundus images to prevent optic nerve damage. A Hybrid Deep Learning Framework for automatic OD and OC segmentation is presented in this work. To efficiently capture pertinent information for glaucoma detection, the model incorporates advanced features like squeeze-and-excitation (SE) blocks, multi-scale attention processes, and Atrous Spatial Pyramid Pooling (ASPP) modules. Experimentation was conducted using the REFUGE dataset for training, and the model was further evaluated on six publicly available fundus image datasets: ORIGA, DRISHTI-GS1, HRF, Dr. HAGIS, BEH, and DRIVE, constituting a cross-dataset assessment to evaluate generalization capability across diverse imaging conditions. The segmentation performance of the proposed model is remarkable, achieving high F1-scores for both OD and OC across multiple datasets. The framework outperformed several state-of-the-art designs, including modified U-Net and linear-dual attention mechanisms, with improvements of up to 6.56% in OC and 2.59% in OD segmentation. The model demonstrated excellent OD segmentation on the DRIVE dataset, achieving an F1-score of 0.9848, highlighting its robustness, accuracy, and versatility across diverse fundus imaging conditions.
The agricultural sector is the most important asset in the country, which boosts development by reducing unemployment, food shortages, poverty, and unstable economic conditions. The recognition and classification of agricultural imagery are the fundamental requisites of contemporary farming methodologies. Thus, it helps to accurately identify the enumeration of plant growth, plant diseases and so on, leading to an increase in crop yield production. There are numerous methods for classifying crop images; however, the conventional methods struggle with issues like noise, distortion, and image quality. It is not capable of fully exploiting the rich potential of image characteristics due to the weather conditions and shooting angles. It does not have the ability to significantly extract the relevant information in the training process and enhance the negative classified outcomes. Multiple learning-based methods have been created to gain more accurate and trustworthy information from image classification. In order to overcome these challenges, a novel crop image classification model is implemented in this research work for classifying the crop images to improve yield production. The required high-quality crop images are collected from the standard datasets. These images are further fed into the feature extraction phase, the Visual Geometry Group 16 (VGG16), graph convolutional neural network (GCNN), and residual network (ResNet) models are utilized to extract the relevant primary information in the collected crop images for generating ensemble convolution features. Subsequently, the resultant features are fed into the classification process, where the multi-scale and dilated adaptive recurrent neural network (MDARNN) mechanism is utilized to effectively classify the crops. In this process, a random function improved dark forest algorithm (RFIDFA) is employed for tuning the RNN parameters, thus improving the classification process. Finally, the validation of the designed approach is performed using several performance measures and compared with the previous works to ensure the designed model’s superior performance. The designed method achieves the best outcome of 95.19% precision, 95.14% NPV, and 95.15% accuracy measures. The developed crop image classification method can continuously monitor the crop growth information to improve yield production. Also, it helps to optimally detect the disease and pest-affected crops in an earlier stage to reduce crop losses, allowing for proper treatment and preventing widespread infection. It is possible to determine the water needs of various crops to ensure efficient crop production and precise farming practices.
Hemoglobin measurement serves as a critical indicator in healthcare, providing information on an individual's physiological well-being and helping diagnose and manage various medical conditions such as anemia and thalassemia. Traditional hemoglobin assessment methods often involve invasive procedures such as blood sampling, which can be uncomfortable, time-consuming, and pose certain risks to patients. In this study, a novel approach is proposed for estimating hemoglobin using distinctive parameters derived from two-wavelength (i.e. 660 and 940nm) PPG signals, in conjunction with ML techniques. To improve the feature extraction process, the PPG signals were decomposed using the downsampling-based averaged wave decomposition technique, yielding a comprehensive set of 68 features from each PPG wave. A composite approach is adopted to optimize feature selection, combining Lasso, Chi-square score, and correlation analysis methods. The resultant set of characteristics, comprising the 10, 20, and 30 point decomposed PPG signals, were systematically evaluated based on their performance in hemoglobin estimation. Data were collected from 150 participants from the Outpatient Block of Thanjavur Hospital. Four regression models - XGBoost, Random Forest, Support Vector Regression, and Logistic Regression were used for this assessment. The findings underscore the superiority of the RFR model utilizing the 30-point decomposed Lasso-selected characteristics, producing an impressive R2 of 0.984, accompanied by a root mean square error (RMSE) of 0.247g/L and a mean absolute error (MAE) of 0.106g/L. This study shows that integrating features from two-wavelength PPG signals with the random forest regression algorithm, along with the use of the Multi-Point Averaged Wave Decomposition Technique, substantially enhances the accuracy of non-invasive hemoglobin measurement in clinical settings.
Muscle fatigue is a neuromuscular condition characterized by a decline in the force-generating capacity of skeletal muscles. Continuous monitoring of this condition plays a crucial role in fields such as sports science, human-machine interface, and ergonomics. Surface electromyography (sEMG) is a widely used non-invasive method for evaluating fatigue conditions. However, the random fluctuations in amplitude and the indeterministic nature of sEMG pose significant challenges for analysis. In this study, an analysis based on slope entropy (SlopEn) is proposed to differentiate the non-fatigue and fatigue conditions under isometric contractions. For this purpose, sEMG signals are acquired from 90 volunteers while they perform an isometric fatiguing task with a 3 kg load in their dominant hand. The first and final two seconds of sEMG are considered to be non-fatigue and fatigue conditions, respectively, and are pre-processed to remove noise. These signals are then subjected to SlopEn to quantify the complexity associated with random fluctuations in the sEMG of muscle contractions. The results show that SlopEn is higher in the fatigue condition, and it is distinct between the two conditions (p < 0.001). Further, SlopEn is found to be superior to Shannon (SE) and Sample entropy (SampEn). The XGBoost achieved an F1 score of 84.1% using both SlopEn and SE. It appears that slope entropy-based models have the potential to detect the fatigue mechanisms in real-time applications.
Hand grip strength (HGS) is a reliable indicator of upper body strength and muscular capacity. This is a comparative study of the various anthropometric variables between college-going students and manual laborers. This study attempts to establish the interrelationship between Hand Grip Strength, Electromyogram, Body Mass Index, age, and fatigue time and lifestyle by examining a group of young college students and a group of young and middle-aged manual laborers. Twenty-five laborers in the age group of 20-60 and 26 college-going students in the age group of 20-30 were randomly selected from BIT Mesra, Ranchi, Jharkhand, India, to assess and compare their HGS and endurance in terms of fatigue time. The HGS and EMG were recorded using the BIOPAC MP45 2-channel data acquisition system along with an SS25LB hand dynamometer (0-50 kg), in EMG-02 Mode of BIOPAC Student Lab Software. BMI shows weak correlation with HGS in students, but the interconnection between higher HGS and dominant arm, as well as male sex, was re-established in this study. No relationship could be established between EMG and HGS. The relationship between HGS and age is strong, with maximum HGS observed in laborers who were in their late 30s and early 40s. While the effect of lifestyle differences didn't reflect in the case of HGS (clench force), as both groups had practically the same HGS, however, in the case of hand grip endurance (fatigue time), the laborers pulled ahead of the students by a significant margin.
This work presents a fully-analog ECG motion artifact elimination circuit that can extract and remove undesired signal components from the ECG waveform corrupted with motion artifacts without using digital processing. The key concept of the proposed scheme is extracting the undesired signal component from the ECG waveform corrupted with motion artifacts by performing moving average, downsampling, and linear interpolation operations. In this case, the huge R, S-peaks that can mislead the extraction process are detected and excluded from the undesired component. Furthermore, the proposed algorithm can be effectively implemented with discrete-time analog blocks, including the high-Q switched-capacitor biquad. As a result, the proposed circuit can be easily incorporated in the ECG sensor frontend, which can lead to a compact and low-power ambulatory ECG monitoring device. The proposed fully-analog ECG motion artifact elimination circuit is realized using CMOS 0.18 & micro;m technology with a core area of 0.85 mm & times; 2.16 mm and power consumption of 6.5 & micro;W. The prototype ECG device, including the proposed circuit, was verified with 22 test participants within the age group from 10 to 65. Test results show the average Percent Root Mean Square Difference (PRD) for the ECG waveforms obtained under various motions (walking, working, and sleeping) compared to steady-state ECG is 21.64% and 5.61% before and after motion artifact elimination.
Photoparoxysmal response analysis in mild cognitive impairment (MCI) and Alzheimer's disease (AD) patients is essential for detecting impaired neural adaptability and dysfunction in response to external stimuli, such as photic stimulation (PS), enabling the early identification of neurocognitive disorder progression. This study investigates photoparoxysmal responses by analyzing PS-induced time-domain and frequency-domain linear features, including skewness, kurtosis, power across frontal, parietal, and occipital regions, power asymmetry, organization score, and cognition index, along with nonlinear brain dynamics using phase-amplitude coupling (PAC) analysis in MCI and AD patients. EEG signals were recorded from patients exposed to repetitive PS frequencies ranging from 3Hz to 30Hz, preprocessed to remove noise artifacts, and analyzed using LASSO regularization to select significant EEG channels. Extracted time-domain and frequency-domain linear features from the power spectrum were used to evaluate oscillatory neural activity and signal variability, while PAC analysis provided insights into nonlinear interhemispheric connectivity and neural adaptability. Results revealed that, with increasing PS frequency, skewness, kurtosis, power asymmetry, and organization scores were lower in MCI than AD, while frontal, parietal, and occipital power, and cognition index were higher in MCI than AD, indicating a transitional shift in brain dynamics associated with disorder progression. Additionally, AD patients exhibited weaker neural responses through low PAC strength, reflecting cortical rigidity and diminished neural flexibility, whereas MCI patients retained more PAC strength, suggesting well interhemispheric connectivity. In AD, delta-alpha and delta-beta coupling strength in the frontal, parietal, and occipital regions was significantly reduced, highlighting progressive dysfunction in PAC modulation. These findings emphasize the importance of linear hand-crafted features with nonlinear PAC representation to effectively assess photoparoxysmal responses in MCI and AD patients, offering a promising approach for the early detection of neurocognitive disorders.
Early detection and diagnosis of brain tumors (BT) increases the likelihood of recovery as well as the number of medical options available to the patient. BT may be detected and diagnosed through magnetic resonance imaging (MRI). Nevertheless, in clinical practice, medical practitioners can only detect brain tumors based on time and experience with a large volume of MRI scans. The computer-aided expert systems are being more utilized to assist in diagnosis and medicine therapy recommendations. A lot of deep learning and machine learning models are applied to identify BT. A Multi-Scale Time-Frequency Convolutional Recurrent Neural Network for Segmentation and Classification of BT has been proposed in this paper (MTF-CNN-SCBT). First, the images are fed using the BraTS Dataset. To implement this, pre-processing is done on the input image by Range-Doppler Matched Filter (RDMF); after which, the processing images are further segregated with Sparsity Fuzzy C-Means Clustering (SFCMC). Ternary Pattern with Discrete Wavelet Transforms (TPDWT) is used to do feature extraction after segmentation. MTF-CNN is then fed with the extracted features to successfully classify BT among Glioma, Meningioma, Pituitary, and No Tumor. Overall, MTF-CNN fails to articulate the adjustment of optimization approaches that can identify the best options to achieve proper brain tumor classification. Therefore, the Coati Optimization Algorithm (COA) is able to optimize MTF-CNN that is able to classify the BT correctly. The suggested MTF-CNN-SCBT is then run in MATLAB, where the performance measures such as F1-score, accuracy, precision, sensitivity and comparison are compared. The MTF-CNN-SCBT proposed model has an absolute accuracy of 94.75% on the BraTS dataset. The proposed MTF-CNN-SCBT model achieves an absolute accuracy of 94.75% on the BraTS dataset. The values 18.75%, 26.89%, and 32.57% represent the percentage improvement of the proposed method over the existing 3D-UNet-DNN-SCBT, SVM-SCBT, and BTC-SAGAN-CHA-MRI methods.
Diabetes Retinopathy (DR) is a serious complication evolving from long-term diabetes mellitus dysfunction affecting the retinal blood vessels. Early diagnosis and medication are crucial to reduce the severity of DR and avert vision loss. However, the existing DR detection and classification techniques fail to extract the subtle lesion features, resulting in limited accuracy. Consequently, this research proposes the Haliaeetus Search and Tempt optimized Self-supervised DCNN-based extreme gradient Boosting (HST-S2CBoost) framework for improving the DR Classification. Specifically, the proposed research exploits the HST-U2Net-based segmentation for performing the effective segmentation of complex lesions such as microaneurysms and exudates from the retinal blood vessels. Besides, the Haliaeetus Search and Tempt optimization (HSTO) enhances the segmentation and optimally tunes the hyperparameters of the classifier, resulting in accurate DR classification. Further, the Statistical Deep Triangular Heat Pattern (SDTH) is employed in the proposed approach for the extraction of deeper structures, leading to improved DR classification. Moreover, the integration of the advanced strategies in the proposed approach contributes to deep feature extraction, reduces the computation complexity, and provides enhanced DR detection and classification. Experimental results demonstrate that the proposed HST-S2CBoost framework achieves superior performance, obtaining a high accuracy of 97.71%, sensitivity of 95.96%, and specificity of 98.70% with 80% of training utilizing the APTOS dataset. Further, the proposed framework achieves superior results of 96.88%,94.92%, and 97.96% for accuracy, sensitivity, and specificity with 80% of training utilizing the IDRiD dataset.
Thyroid cancer is a form of cancer that begins in the thyroid gland. Precise and efficient treatments can be provided for thyroid cancer if nodules are detected early, thereby greatly lowering morbidity and mortality. Early thyroid nodule detection has been achieved by applying the traditional approach based on ultrasound imaging widely. A Deep Learning (DL)-based technique for classifying thyroid cancer is introduced here, as the conventional methods are time-consuming, expensive, and sometimes even ineffective. Hence, a Hybrid Residual Zeiler and Fergus Network (HyRZNet) framework is developed in this research for thyroid cancer classification. Initially, thyroid ultrasound images are obtained from the specified database and are subjected to image pre-processing, where a Bilateral Filter (BF) is employed to eradicate the noise in the image. Later, the pre-processed image is given to cancer-affected region detection using a Mask Region-based Convolutional Neural Network (Mask RCNN). After cancer region detection, a feature extraction process is carried out, where texture features, namely, Complete Local Binary Pattern (CLBP), Local Vector Pattern (LVP), and Grey-level Co-occurrence Matrix (GLCM) features are extracted. Finally, thyroid cancer classification is carried out by HyRZNet and the accuracy obtained is 90.80%, True Positive Rate (TPR) is 92.46%, and True Negative Rate (TNR) is 89.20%. Compared to the existing classifiers, the proposed classifier is more accurate.
Fine-tuning involves customizing a pre-trained Large Language Model (LLM) for a particular domain like medicine by retraining it on a smaller specialized dataset. This process helps the LLM to effectively manage complex data such as lengthy Electronic Health Records (EHRs) discharge summaries, task difficult for humans and traditional methods. This challenge was specifically significant in rapidly-evolving medical field. An LLM with outdated knowledge can generate inaccurate or fabricated information in high-stakes healthcare situations. This problem is compounded for non-English languages, which often lacks necessary benchmarking tools. In order to overcome these problems in the traditional method, a novel Fine Tuned LLMs (FTLLMs) method for a clinical question-answering model is implemented in this proposed system. At the earlier stage, the text essential for the evaluation is fetched from the benchmark resources. Further, these fetched texts are subjected to text pre-processing. In the pre-processing work, four different tasks are computed such as removal of punctuation and special characters, elimination of redundant and irrelevant data, removal of stop words, and stemming. Next, pre-processed text is fed for the prediction of the medical QA system using the proposed FTLLM. Moreover, parameters are fine-tuned by newly introduced Fitness Sorted-based Artificial Protozoa Optimizer (FSAPO). Through optimization process, performance metrics like F1-score and the exact match were maximized. In addition, several functionality calculations are improved in the proposed Fine Tuned LLM (FTLLM) by optimizing the network for a medical QA system through existing methods. The newly designed FTLLM fine-optimizing model for the medical QA system is introduced in Python, and the detection process is also performed. Here, the superiority of the implemented model is contrasted with the previous systems and with the statistical metrics. A comparative analysis of Dataset 1 demonstrates that the proposed FSAPO-FTLLM algorithm, when utilizing the ReLU activation function achieves 91% increase in precision compared to existing performance baseline techniques. This ReLU activation function reduces 94.7% computational load by using threshold operation and provides faster performance training performance.
