Scholarlink

In severe cases, COVID-19 progresses to pneumonia, leading to suffocation, breathlessness, and, in extreme instances, multi-organ failure and death. In response to these challenges, extracting and analyzing lung patterns can be beneficial through a deep neural network (DNN) based lung segmentation. The proposed DNN comprises numerous distinct operations. The filter size of the 3 × 3 convolution operation followed by ReLU extracts feature maps of large images in individual tiles. The max pooling operation reduces the image size of the feature map. It propagates the maximum activation from each 2 × 2 window to the next feature map and increases feature channels by a factor of two. It results in a spatial contraction and assigns all features into a singular output vector. In the expansion pathway, the upconvolution operation maps each feature vector to a 2 × 2 output window followed by 3 × 3 convolution and ReLU. It concatenates with the corresponding high-resolution features obtained from the contracting pathway. The resulting segmentation map comprises two channels designated for foreground and background classes. The image size of the feature map varies from 512 × 512 to 32 × 32 and vice versa. The feature channel (no. of filter) ranges from 32 to 512 and vice versa. The proposed model is analyzed with different values of parameters like kernel or filter size, number of filter, pooling strategies, optimizer, learning rate, activation function, loss function, batch size and epochs. The best hyperparameter values are obtained, and training procedures are improved in the proposed modified UNet. By leveraging this advanced architecture, the segmentation process benefits from extracting low-level and high-level properties using encoders, skip connections and decoders. It combines data from different resolutions to emphasize critical areas. Our results, based on publicly accepted image datasets of chest X-rays and CT-scans from Kaggle, demonstrate the effectiveness of our proposed method. The metrics such as accuracy and dice coefficient are used for quantitative evaluation. The proposed lung segmentation model's training, validation and test accuracy is 95%, 94.8% and 92.1% for X-ray images and 95.1%, 94.4%, and 91.2% for CT-scan images. The performance of the proposed model is comparable with existing models and can used by clinicians in diagnosing and monitoring COVID-19 cases.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lung-UNet: A Modified UNet-Based DNN for COVID Lung Segmentation from Chest X-Ray and CT-Scan Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sanjib Saha

摘要

In severe cases, COVID-19 progresses to pneumonia, leading to suffocation, breathlessness, and, in extreme instances, multi-organ failure and death. In response to these challenges, extracting and analyzing lung patterns can be beneficial through a deep neural network (DNN) based lung segmentation. The proposed DNN comprises numerous distinct operations. The filter size of the 3 × 3 convolution operation followed by ReLU extracts feature maps of large images in individual tiles. The max pooling operation reduces the image size of the feature map. It propagates the maximum activation from each 2 × 2 window to the next feature map and increases feature channels by a factor of two. It results in a spatial contraction and assigns all features into a singular output vector. In the expansion pathway, the upconvolution operation maps each feature vector to a 2 × 2 output window followed by 3 × 3 convolution and ReLU. It concatenates with the corresponding high-resolution features obtained from the contracting pathway. The resulting segmentation map comprises two channels designated for foreground and background classes. The image size of the feature map varies from 512 × 512 to 32 × 32 and vice versa. The feature channel (no. of filter) ranges from 32 to 512 and vice versa. The proposed model is analyzed with different values of parameters like kernel or filter size, number of filter, pooling strategies, optimizer, learning rate, activation function, loss function, batch size and epochs. The best hyperparameter values are obtained, and training procedures are improved in the proposed modified UNet. By leveraging this advanced architecture, the segmentation process benefits from extracting low-level and high-level properties using encoders, skip connections and decoders. It combines data from different resolutions to emphasize critical areas. Our results, based on publicly accepted image datasets of chest X-rays and CT-scans from Kaggle, demonstrate the effectiveness of our proposed method. The metrics such as accuracy and dice coefficient are used for quantitative evaluation. The proposed lung segmentation model's training, validation and test accuracy is 95%, 94.8% and 92.1% for X-ray images and 95.1%, 94.4%, and 91.2% for CT-scan images. The performance of the proposed model is comparable with existing models and can used by clinicians in diagnosing and monitoring COVID-19 cases.