Search

• To propose a novel hybridized deep learning technique (HyKBRNet-DLSTM) for automatic detection of varicose vein (V2) with the aid of leg vein features. • To pre-process the raw images using Median Kuan Filtering (Me_KF) technique for enhancing the visual quality of the image. • To extract the vein related features using Extended Twofold Autoencoder (ExT_AE) technique for enhancing the accuracy performance. • To perform dimensionality reduction using Tent Chaotic Zebra Optimizer (TCh-ZO) algorithm for reducing data redundancy and overfitting issues in the proposed model. • To detect the presence and absence of varicose veins with the use of novel hybridized Kernel Boosted ResNet-Dropped Long Short Term Memory (HyKBRNet-DLSTM) technique. Varicose Veins (V2) are the most harmful chronic disease when blood circulation of leg veins is not working precisely. This results in serious blood circulation problems from the heart to the leg. Diagnosing the V2 problem is highly cost-effective; earlier detection is necessary for the patients to relieve pain effectively. Hence, this article brings a novel hybridized deep learning (DL) technique for efficiently detecting the presence/absence of V2. The Median Kuan Filtering (Me_KF) based pre-processing technique is introduced to filter the noises from the raw leg vein images. After pre-processing, the Extended Twofold Autoencoder (ExT_AE) based DL technique is proposed to extract leg vein features. Then, Tent Chaotic Zebra Optimizer (TCh-ZO) algorithm is proposed to eliminate redundant features and enhance accuracy performance. Finally, this study introduces the Hybridized Kernel Boosted ResNet-Dropped Long Short Term Memory (HyKBRNet-DLSTM) based DL technique to detect the presence and absence of varicose veins. The experiment is processed with 11,350 leg vein images, and measures like accuracy, precision, kappa, mean square error (MSE) and time complexity are analyzed. The experimental section proposes an accuracy of 98.5%, a precision of 98.69%, a kappa of 95.77%, an MSE of 1.73 and a time complexity of 25.36 ms. [ABSTRACT FROM AUTHOR]

Highlights • Feature sets robust against force level variations in pattern recognition based control of myoelectric upper limb prosthesis have been proposed. • Significance of using non-linear techniques to analyze complex signals like EMG, which are both non-linear and non-stationary has been highlighted here. • With a significant improvement in classification performance of about 7% to 9%, the proposed feature sets, FFS and PSEAR outperformed all the other well established feature sets for contraction level invariant myoelectric control, considered in this paper. • Improvement in performance was observed not only for flexion movements but also for prehension and power grip movements. • The proposed feature sets can be potential candidates to make the clinical implementation of intuitive PR-based myoelectric prosthetic control possible in the future. Abstract In spite of the tremendous progress of upper limb myoelectric prosthetic control in the field of rehabilitation engineering, there still exist several real world challenges to be met, before realizing it as a good substitute for a natural arm. Incompetence of the system to accommodate variations in contraction levels of muscle movements has been identified as one of the significant challenges, as these variations have a subsequent impact on the performance of pattern recognition based myoelectric control. Non-linear techniques are more suited to characterize myoelectric signals since one of their major properties is nonlinearity. Based on this we propose two feature combinations which can lead to a reliable control scheme that is robust against contraction level variations. The performance of our proposed features when tested on nine transradial amputees for six motion classes at three different force levels outweighed other established feature extraction methods meant for contraction variation independent control. Significant improvement of around 8% in average classification performance was achieved across all subjects and force levels, subjected to training, both with all force levels and with unseen force levels. Moreover, these features achieved superior performance in classifying flexion as well as grip movements. [ABSTRACT FROM AUTHOR]