http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
A Novel Integrated Convolutional Neural Network via Deep Transfer Learning in Colorectal Images
Sahadev Poudel,이상웅 조선대학교 IT연구소 2019 정보기술융합공학논문지 Vol.9 No.2
In this paper, we explore the use of current deep learning methods, convolutional neural networks (CNNs) in the field of computer-aided diagnosis systems to classify several endoscopic colon diseases. Transfer learning by fine-tuning deep convolutional neural networks (CNNs) is applied due to the limited amount of data. For this, state-of-the-art CNN architectures, such as VGG16, VGG19, InceptionV3, ResNet50, Inception-ResNet-V2, DenseNet169 were used for training and validating the dataset. However, these existing architectures cannot extract more dense endoscopic image features and have problem on similar-looking images of different category. Therefore, we propose a novel integrated convolutional neural network to develop a more accurate and highly efficient method for endoscopic image classification, which uses the features of earlier layers in the classification process and increases the receptive field of view at the end layers in the network. We compare and evaluate our performance using performance metrics Accuracy (ACC), Recall, Precision and F1-score. In our experimental results, the proposed method outperforms the existing architectures, obtaining an accuracy of about 92.4% on the test dataset.
Polyp Segmentation and Generalization using Style Conversion
Astha Adhikari,Sahadev Poudel,Sang-Woong Lee 한국차세대컴퓨팅학회 2022 한국차세대컴퓨팅학회 학술대회 Vol.2022 No.10
In recent years, the deployment of deep neural networks in real-time clinical settings has been considered vulnerable due to domain shifts, which lowers their performance. Polyp image has significant appearance shifts, which eventually impact the performance. So, a deep learning model that generalizes unseen images is in high demand. This paper introduced a practical approach to improving domain shift issues. Firstly, we unified the style transfer with the segmentation model into one framework to diminish the appearance shifts problems and do segmentation alongside. Secondly, with the help of Adaptive Instance Normalization, we transferred the style precisely and dynamically in the earlier layers of the segmentation model. Our solution shows better results on the 224*224 image input than other baseline models.
Raman Ghimire,Sahadev Poudel,Sang-Woong Lee 한국차세대컴퓨팅학회 2021 한국차세대컴퓨팅학회 학술대회 Vol.2021 No.05
In recent years, UNet architecture has shown to be a standard network for medical image segmentation. However, it suffers from some severe limitations. It loses localization ability for low-level details followed by the inability of long-range dependencies. Motivated by this, we explore transformer-based architectures that exploit global context by modeling long-range spatial dependencies, which are essential for accurate polyp segmentation. In this paper, we propose an attention-based transformer encoded UNet model. This hybrid model inherits both characteristics of CNN block as well as attention block. We perform various experiments in existing architectures like UNet, ResUNet, ResUNet-Mod and our proposed method. The proposed method achieved a 0.645 mIOU score took an unassailable lead over prior methods.