粗读Deep Learning COVID-19 Features on CXR Using Limited Training Data Sets

面对有限的数据集，本文提出了一种新的训练方法，通过分割，找出病灶可能存在的区域，然后针对这个区域，随机多次采样，进行分类。

首先，针对简单预处理的肺部X光图像，分割出肺部区域，并利用mask抹去非肺部区域的像素。

之后的分类网络才是重点。大多数方法把整个肺部图像缩小到224*224，然后直接塞给网络进行训练。而本文则是进行裁剪采样，每次从确定包含肺部的区域裁处224*224大小的图像块，然后把整幅图像的标签赋予这个裁剪得到的图像块，并让网络进行学习。

在推断阶段，模型会从每一幅测试图像中采样K个图像块，作者令K=100，这100各块都输入刚刚训练的Resnet18中进行分类，最后以这K个块的分类结果中，出现最多的那一种作为模型认定的分类结果。

粗读Deep Learning COVID-19 Features on CXR Using Limited Training Data Sets相关推荐

《学习geometric deep learning笔记系列》第一篇，Non-Euclidean Structure Data之我见
<学习geometric deep learning笔记系列>第一篇,Non-Euclidean Structure Data之我见 FesianXu at UESTC 前言本文是笔者在 ...
【论文阅读】Deep Compositional Captioning: Describing Novel Object Categories without Paired Training Data
[论文阅读]Deep Compositional Captioning: Describing Novel Object Categories without Paired Training Data ...
[论文笔记 ECCV2020] Learning to Count in the Crowd from Limited Labeled Data
[论文笔记 ECCV2020] Learning to Count in the Crowd from Limited Labeled Data 摘要 Abstract 贡献 Contribution ...
[文献阅读] Sparsity in Deep Learning: Pruning and growth for efficient inference and training in NN
文章目录 1. 前言 2. Overview of Sparsity in Deep Learning 2.1 Generalization 2.2 performance and model sto ...
Towards Deep Learning Models Resistant to Adversarial Attacks(PGD adversarial training)
目录 Introduction 内容简介 An Optimization View on Adversarial Robustness 内容介绍 Towards Universally Robust ...
粗读SELF-SUPERVISED LEARNING IS MORE ROBUST TO DATASET IMBALANCE
本文首先通过两组实验,证明自监督预训练能够更好地应对极端的数据不平衡现象 ID是in-domain缩写,而OOD是out-of-domain缩写.前者代表预训练数据集和测试集是同一数据集,而后者说明两 ...
【Deep Learning】笔记：Understanding the difficulty of training deep feedforward neural networks
这几天读了这篇论文,在这里将大致内容写在这里. Abstract 介绍这篇论文的主要内容就是尝试更好的理解为什么使用"标准随机初始化"来计算使用标准梯度下降的网络效果通常来讲都不是 ...
【论文笔记】Learning to Count in the Crowd from Limited Labeled Data
文章目录 Abstract 1 Introduction 3 Preliminaries 4 GP-based iterative learning 4.1 Labeled stage 4.2 Unl ...
Automated defect inspection system for metal surfaces based on deep learning and data augmentation
Automated defect inspection system for metal surfaces based on deep learning and data augmentation 基 ...

粗读Deep Learning COVID-19 Features on CXR Using Limited Training Data Sets

粗读Deep Learning COVID-19 Features on CXR Using Limited Training Data Sets相关推荐

最新文章

热门文章