Lightweight Model For The Prediction of COVID-19 Through The Detection And Segmentation of Lesions in Chest CT Scans

We introduce a lightweight Mask R-CNN model that segments areas with the Ground Glass Opacity and Consolidationin chest CT scans. The model uses truncated ResNet18 and ResNet34 nets with a single layer of Feature Pyramid Networkas a backbone net, thus substantially reducing the number of the parameters and the training time compared to similarsolutions using deeper networks.Without any data balancing and manipulations, and using only a small fraction ofthe training data, COVID-CT-Mask-Net classification model with 6.12M total and 600K trainable parameters derivedfrom Mask R-CNN, achieves 91.35% COVID-19 sensitivity, 91.63% Common Pneumonia sensitivity, 96.98% true negativerate and 93.95% overall accuracy on COVIDx-CT dataset (21191 images). We also present a thorough analysis of theregional features critical to the correct classification of the image. The full source code, models and pretrained weightsare available on https://github.com/AlexTS1980/COVID-CT-Mask-Net.

通过胸部CT扫描中病变的检测和分割来预测COVID-19的轻型模型

我们引入了一种轻量级的Mask R-CNN模型，该模型在胸部CT扫描中使用毛玻璃的不透明度和合并来分割区域。该模型使用具有单层特征金字塔网络的截断的ResNet18和ResNet34网络作为骨干网，从而与使用更深层网络的类似解决方案相比，大大减少了参数数量和训练时间。.. 在没有任何数据平衡和操纵的情况下，仅使用训练数据的一小部分，COVID-CT-Mask-Net分类模型具有总计6.12M的可训练参数和来自Mask R-CNN的600K可训练参数，可实现91.35％的COVID-19灵敏度，在COVIDx-CT数据集（21191张图像）上，普通肺炎的敏感性为91.63％，真实阴性率为96.98％，总体准确性为93.95％。我们还将对对图像正确分类至关重要的区域特征进行全面分析。完整的源代码，模型和预训练权重可在https://github.com/AlexTS1980/COVID-CT-Mask-Net上获得。 （阅读更多）