Wyniki wyszukiwania - BazTech

1

Cell image augmentation for classification task using GANs on Pap smear dataset

Zak Jakub, Grzeszczyk Michal K., Pater Antonina, Roszkowiak Lukasz, Siemion Krzysztof, Korzynska Anna

Biocybernetics and Biomedical Engineering

|

2022

|

Vol. 42, no. 3

995--1011

EN

One of the solutions to the problem of insufficiently large training datasets in image processing is data augmentation. This process artificially extends the size of training datasets to avoid overfitting. Generative Adversarial Networks yield that become increasingly difficult to differentiate from real images, until the differentiation is no longer possible. Thus, artificial images closely resembling original ones can be generated. Inclusion of artificial images contributes to improving the training process. Medical domain is one of the areas where data acquisition is burdened by many procedures, laws, and prohibitions. As a result the potential size of collected datasets is reduced. This article presents the results of training Convolutional Neural Networks on a artificially extended image datasets. The resulting classification accuracy on a cell classification task of models trained with images generated using the proposed method were increased by up to 12.9% in comparison to that of the model trained only with original dataset from the HErlev Pap smear dataset.

2

Variational mode decomposition-based seizure classification using Bayesian regularized shallow neural network

Yadav Vipin Prakash, Sharma Kamlesh Kumar

Biocybernetics and Biomedical Engineering

|

2021

|

Vol. 41, no. 2

402--418

EN

This work presents a new epileptic seizures epoch classification scheme. Variational mode decomposition (VMD), has been explored for non-recursively decomposing the electroencephalogram (EEG) signals into fourteen band limited intrinsic mode functions (IMFs). Data augmentation (DA), has been used for handling unbalanced classification problem. Normalized energy, fractal dimension, number of peaks, and prominence parameters were computed from the band-limited IMFs for the discrimination of seizure and non-seizure epochs. Bayesian regularized shallow neural network (BR-SNNs) and six other well-known classifiers were tested. Sensitivity, specificity, and accuracy have been used as performance metrics. This study includes two different epoch lengths of 1-second and 2-seconds. A total of 32 test cases for both, class balanced and unbalanced classification problems have been taken for the performance evaluation. The best performance obtained is 100% for all the three metrics from the test cases of database-2 and 3. For database-1, average sensitivity, specificity, and accuracy of 99.71, 99.75, and 99.73% have been achieved, respectively for the 1-second epoch. The presented work shows better performance results compared to many previously reported works.

3

Classification of lung nodule malignancy in computed tomography imaging utilising generative adversarial networks and semi-supervised transfer learning

Apostolopoulos Ioannis D., Papathanasiou Nikolaos D., Panayiotakis George S.

Biocybernetics and Biomedical Engineering

|

2021

|

Vol. 41, no. 4

1243--1257

EN

The pulmonary nodules’ malignancy rating is commonly confined in patient follow-up; examining the nodule’s activity is estimated with the Positron Emission Tomography (PET) system or biopsy. However, these strategies are usually after the initial detection of the malignant nodules acquired from the Computed Tomography (CT) scan. In this study, a Deep Learning methodology to address the challenge of the automatic characterisation of Solitary Pulmonary Nodules (SPN) detected in CT scans is proposed. The research methodology is based on Convolutional Neural Networks, which have proven to be excellent automatic feature extractors for medical images. The publicly available CT dataset, called Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI), and a small CT scan dataset derived from a PET/CT system, is considered the classification target. New, realistic nodule representations are generated employing Deep Convolutional Generative Adversarial Networks to circumvent the shortage of large-scale data to train robust CNNs. Besides, a hierarchical CNN called Feature Fusion VGG19 (FF-VGG19) was developed to enhance feature extraction of the CNN proposed by the Visual Geometry Group (VGG). Moreover, the generated nodule images are separated into two classes by utilising a semi-supervised approach, called self-training, to tackle weak labelling due to DC-GAN inefficiencies. The DC-GAN can generate realistic SPNs, as the experts could only distinguish 23% of the synthetic nodule images. As a result, the classification accuracy of FF-VGG19 on the LIDCIDRI dataset increases by +7%, reaching 92.07%, while the classification accuracy on the CT dataset is increased by 5%, reaching 84,3%.

4

Application of residual learning to microseismic random noise attenuation

Zheng Jing, Jiang Tianqi, Wu Zhixiang, Sun Yuan

Acta Geophysica

|

2021

|

Vol. 69, no. 4

1151--1161

EN

Microseismic data which are recorded by near-surface sensors are usually drawn in strong random noise. The reliability and accuracy of arrivals picking, source localization, microseismic imaging and source mechanism inversion are often afected by the random noise. Random noise attenuation is important for microseismic data processing. We introduce a novel deep convolutional neural network-based denoising approach to attenuate random noise from 1D microseismic data. The approach predicts the noise (the diference between the noisy microseismic data and clean microseismic data) as output instead of directly outputing the denoised data that is called residual learning. With the residual learning strategy, the approach removes the clean data in the hidden layers. In other words, the approach learns from the random noise prior instead of an explicit data prior. Then, the denoised data are reconstructed via subtracting noise from noisy data. Compared with other commonly used denoising methods, the proposed method performs its efectiveness and superiority by experimental tests on synthetic and real data. The model is trained with synthetic data and applied on real data. The results show that random noise in the synthetic and real data can been removed. However, some noise still remains in the real data case. The reason for that may be the approach can only remove random noise nor the correlated noise. Other methods are needed to be applied to remove the correlated noise to obtain higher performance after that approach when the real microseismic data which contain both correlated noise and random noise.

5

A deep Residual U-Net convolutional neural network for automated lung segmentation in computed tomography images

Khanna Anita, Londhe Narendra D., Gupta S., Semwal Ashish

Biocybernetics and Biomedical Engineering

|

2020

|

Vol. 40, no. 3

1314--1327

EN

To improve the early diagnosis and treatment of lung diseases automated lung segmentation from CT images is a crucial task for clinical decision. The segmentation of the lung region from the CT scans is a very challenging task due to the irregular shape and size of lungs, low contrast and fuzzy boundaries of the lung. The manual segmentation of lung CT images is a laborious task. Therefore, various approaches are suggested by the researcher in the recent past for the automated lung segmentation. However, the existing approaches either utilize low-level handcraft features or CNN based Fully Convolutional Networks. The low-level hand-craft feature-based approaches lead to poor generalization, while the shallower networks are unable to extract more discriminative features. Hence, in this study, we have implemented a deep learning-based architecture called Residual U-Net with a false-positive removal algorithm for lung CT segmentation. Here, we have suggested that learning from a substantially deeper network with residual units can extract more discriminative feature representation as compared to shallow network for lung segmentation. To take full advantage of the deeper network, we have utilized a set of schemes to ensure efficient training. First, we have implemented a U-Net architecture with residual block to overcome the problem of performance degradation. Further, various data augmentation techniques are utilized to improve the generalization capability of the proposed method. The experimental results show that the proposed method achieved competitive results over the existing methods with DSC of 98.63%, 99.62% and 98.68% for LUNA16, VESSEL12 and HUG-ILD dataset respectively.