Volume 1 Issue 1, April 2023
Breast Tumor Classification Using Hybrid architectures (Googlenet, Alexnet, VGG16)
Authors: Sri Geetha M
Abstract: Breast cancer is a kind of cancer that develops in the breast cells. Segmentation of breast tumor is a critical task. Segmenting tiny tumours in ultrasound images may be difficult owing to speckle noise, the fact that tumour forms and sizes might differ across individuals, and the presence of imaging areas that resemble tumours. In the area of biomedical image analysis, advanced learning-based algorithms have recently experienced a lot of success; nevertheless, one of most modern approaches that are now accessible have a poor track record whenever it relates to properly separating small breast cancers. This article presents a novel method to detect breast cancer at early stage. The proposed method combines the three architectures namely alexnet, googlenet and VGG. The total number of images utilized in this research came from two sources: small scale MIAS gave 576 images, and INbreast contributed 1095 images. According to the findings, the strategy that was presented gets the best overall performance and exceeds every other approach when it comes to the segmentation of tiny tumours. We aim to develop an approach to the problem of the (FDC) Feature Dimensionality Curve for the profound highlights that are obtained through the exchange learning pre-prepared CNNs. This structure is dependent on having previously created deep convolutional systems in addition to having a worldview that is based on univariate-based huge information. While removing shallow and profound aspects from INbreast mammograms, the deep – learning systems AlexNet, VGG, as well as GoogleNet are arbitrary choices that are employed. In terms of accuracy, loss rate, and runtime, GoogleNet outperformed AlexNet, PSO-MLP, and ACO-MLP, achieving 99% accuracy, the lowest loss rate (0.1557), and the shortest runtime (4.14 minutes).This performance seems to be advantageous for developing a computer-aided diagnostic (CAD) framework that is both practical and trustworthy for breast tumour categorization.
Keywords: deep transfer learning; breast ultrasound; feature reduction and selection;deep learning; feature dimensionality curse (FDC); tumor segmentation; CAD system; small tumor-aware network breast tumor
Extended Kalman filter to detect Human body movements through inertial and magnetic sensors – Medical Applications
Authors: T. Senthil Kumar, Leo John Baptist Andrews
Abstract:
Inertial orientation tracking is something that is being looked into right now, especially when it comes to capturing human movement outside in real time. Quaternions are used to show how the navigation frame and the body frame are connected in a way that makes them rotate together. In this article, we explain how to use a quaternion-based Extended Kalman Filter to figure out how a rigid body is oriented in three dimensions (EKF). The EKF uses the readings from an Inertial Measurement Unit (IMU), which is a combination of a tri-axial magnetic sensor and an accelerometer. The solution that has been suggested is a universal filter that doesn’t set the amount of freedom at the connections among the different parts of the model. Even when the observatory conditions are important, the algorithm’s performance can be measured with computer simulations and real-world testing.
Keywords: Extended Kalman filter, inertia, magnetic sensing, orientation
Automated detection of ulcers and erosions in capsule endoscopy images using a convolutional neural network
Authors: Sivakumar. P
Abstract:
The wireless capsule endoscope, often known as WCE, has been successfully tested and validated in clinical settings for so many years. This new study approach, which is based on artificial intelligence technology and has excellent levels of accuracy and resilience, has the potential to minimise detection costs and help the general population. Furthermore, automatic detection techniques that look for signs of gastrointestinal sickness in WCE pictures have been accepted by these professionals as an ideal enhancing tool. Within the scope of this study, a novel computer-aided diagnostic approach for ulcer identification in WCE pictures is put up as a potential solution. The method was evaluated by an independent party with 256 591 new pictures. When implemented to the validation data, the learning algorithm had a diagnosis accuracy of 98.067 percent. According to the results of our correlations of F1, F2, and ROC-AUC, the proposed approach performs good than a number of off-the-shelf CNN models, such as VGG, Inception-ResNet-v2 and DenseNet, as well as traditional machine learning methods for WCE image classification that involve handcrafted features. Overall, the results of this research show that it is possible to recognise ulcers in WCE photos by using the deep CNN approach. In addition, the HAnet architecture that we have developed specifically for this issue provides an excellent option for the construction of the network structure. The CNN networks that were employed in this study were pre-trained using millions of annotated natural photos in order to compensate for the small number of images of ulcers that were included in the accessible data sets (ImageNet). Following the extraction of the deep features, a randomized forests classifier is used to determine whether or not WCE pictures include ulcers. The outcomes are quite encouraging (96.73 percent and 95.34 percent in terms of precisions and recall accordingly). Keywords: Detection network, gastric ulcer, wireless capsule endoscope, convolutional neural network, complications.
Download the Article
Classification of Body Position during Prayer using the Convolutional Neural Network
Authors: M. Nalini, T.J.Nagalakshmi
Abstract: A Muslim must perform Salat (prayer) five times a day as the most fundamental and important form of religious devotion, as it is the second pillar of Islam. EEG recordings of brain activity during a Namaz can be used to study the effects of rapid changes in body position and a 14- channel EEG recorder monitors the brain activity of 40 Muslim participants during a four-cycle Namaz. Different Namaz positions were used to measure brain connectivity in several frequency bands. An artificial intelligence-assisted framework to assist worshippers in assessing the accuracy of their prayer postures is one solution to these problems. Using Convolutional Neural Networks to recognise basic Islamic prayer movements is the first step in achieving this goal. A YOLOv3 neural network was trained on a dataset of Salat positions to recognise the gestures in this paper. According to the experimental results, for a training dataset of 764 photos.
Keywords: Body position; Namaz; convolutional neural network; cross correlation images
Download the article
Optimization techniques and hybrid deep learning approaches for UV index predictions
Authors: A.Rajasekar, B.R. Tapas Bapu, N.Ashokkumar
Abstract: The sun ultra – violet indices, often known as the UVI, is an important global health statistic that may help reduce the risk of illnesses caused by ultraviolet radiation. For the purpose of forecasting the everyday UVI in India, this research attempted to build as well as analyze the outputs of several hybridised algorithms for machine learning using a neural network convolutional as well as a longer selective memory termed to as CLSTM. The results demonstrated that the recommended modified CEEMDAN-CLSTM concept possesses an excellent capacity for predicting, i.e., a low error rate and an excellent performance level. This was illustrated in comparison to the standard models that were used as the counterparts. The implication of the work has the potential to improve real-time exposure guidance for the general population as well as to assist in mitigating the risk of illnesses associated to solar UV radiation, including melanoma
Keywords: mathematical formulation, optimization method, machine learning prediction of the ultraviolet rays
Download the article
Mathematical Modelling of Multi-Task Cascaded Convolutional Networks: Face Detection Application
Authors: S.B. Mohan
Abstract: In a setting with little restrictions, it might be difficult to identify and align faces because to the wide variety of possible postures, illuminations, and occlusions. Recent research has shown that deep learning methods are capable of achieving outstanding results in both of these categories of problems. In this research, we provide a deep pipelined multi-task structure that, in order to improve the performance of the tasks, takes use of the intrinsic connection that exists between them. In specifically, our approach utilises a cascaded structure consisting of three stages of meticulously crafted deep convolutional networks. This strategy has the ability to enhance effectiveness automatically and does not need any human sample selection. While maintaining real-time performance, our approach achieves higher accuracy than the current state-of-the-art methods on the difficult FDDB as well as WIDER FACE benchmarks for face detection, as well as the AFLW standard for face alignment.
Keywords: identification of faces, mathematical programming, and convolutional neural networks