Abstract

COVID-19—A pandemic declared by WHO in 2019 has spread worldwide, leading to many infections and deaths. The disease is fatal, and the patient develops symptoms within 14 days of the window. Diagnosis based on CT scans involves rapid and accurate detection of symptoms, and much work has already been done on segmenting infections in CT scans. However, the existing work on infection segmentation must be more efficient to segment the infection area. Therefore, this work proposes an automatic Deep Learning based model using Transfer Learning and Hierarchical techniques to segment COVID-19 infections. The proposed architecture, Transfer Learning with Hierarchical Segmentation Network (TLH-Net), comprises two encoder–decoder architectures connected in series. The encoder–decoder architecture is similar to the U-Net except for the modified 2D convolutional block, attention block and spectral pooling. In TLH-Net, the first part segments the lung contour from the CT scan slices, and the second part generates the infection mask from the lung contour maps. The model trains with the loss function TV_bin, penalizing False-Negative and False-Positive predictions. The model achieves a Dice Coefficient of 98.87% for Lung Segmentation and 86% for Infection Segmentation. The model was also tested with the unseen dataset and has achieved a 56% Dice value.

中文翻译：

---

迁移学习 - COVID CT 扫描的分层分割

摘要

COVID-19——世界卫生组织于 2019 年宣布的大流行病已在全球蔓延，导致许多人感染和死亡。这种疾病是致命的，患者会在窗口期后 14 天内出现症状。基于CT扫描的诊断需要快速、准确地检测症状，并且在CT扫描中分割感染方面已经做了很多工作。然而，现有的感染分割工作必须更有效地分割感染区域。因此，这项工作提出了一种基于自动深度学习的模型，使用迁移学习和分层技术来分割 COVID-19 感染。所提出的架构，分层分割网络的迁移学习（TLH-Net），由两个串联的编码器-解码器架构组成。除了修改后的 2D 卷积块、注意力块和谱池之外，编码器-解码器架构与 U-Net 类似。在 TLH-Net 中，第一部分从 CT 扫描切片中分割肺部轮廓，第二部分从肺部轮廓图生成感染掩模。该模型使用损失函数 TV_bin 进行训练，惩罚假阴性和假阳性预测。该模型的肺分割 Dice 系数为 98.87%，感染分割 Dice 系数为 86%。该模型还使用未见过的数据集进行了测试，并取得了 56% 的 Dice 值。