CheXNet for Binary & Multi-Class Pneumonia Detection

Sik-Ho Tsang

·

Follow

3 min read

·

Dec 23, 2023

--

1

Listen

Share

CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning
CheXNet, by Stanford University
2017 arXiv v3, Over 2600 Citations (Sik-Ho Tsang @ Medium)

Biomedical/Medical Image Classification
2017 [ChestX-ray8] 2019 [CheXpert] 2020 [VGGNet for COVID-19] [Dermatology] [Deep-COVID] [Zeimarani ACCESS’20] [Multiview CNN] 2021 [CheXternal] [CheXtransfer] [CheXbreak] 2022 [BUS-CNN] [CheXED]
==== My Other Paper Readings Are Also Over Here ====

• CheXNet, is a 121-layer convolutional neural network (CNN) trained on ChestX-ray14, which contains over 100,000 frontal-view X-ray images with 14 diseases.

Outline

1. CheXNet
2. Results

1. CheXNet

1.1. Dataset

• ChestX-ray14 dataset released by Wang et al. (2017) which contains 112,120 frontal-view X-ray images of 30,805 unique patients.
• Wang et al. (2017) is the author group of the dataset ChestX-ray8.
• A test set of 420 frontal chest X-rays is used.
• The radiologists had 4, 7, 25, and 28 years of experience.

1.2. Binary Classification

• The pneumonia detection task is a binary classification problem, where the input is a frontal-view chest X-ray image X and the output is a binary label y, indicating the absence or presence of pneumonia.
• ImageNet-pretrained 121-layer DenseNet is used.
• Before inputting the images into the network, the images are downscaled to 224×224 and normalized using ImageNet’s mean and standard deviation.
• Random horizontal flipping is applied for data augmentation.
• The final fully connected layer is replaced with one that has a single output with sigmoid nonlinearity applied.
• The weighted binary cross entropy loss is:

• The dataset is randomly split into training (28744 patients, 98637 images), validation (1672 patients, 6351 images), and test (389 patients, 420 images). There is no patient overlap between the sets.

1.3. 14-Pathology Classification

• In this case, CheXNet outputs a vector t=14 of binary labels indicating the absence or presence of each of the following 14 pathology classes.
• The final fully connected layer in CheXNet is replaced with a fully connected layer producing a 14-dimensional output, after which an element-wise sigmoid nonlinearity is applied.
• The sum of unweighted binary cross entropy losses is used:

• The dataset is split into training (70%), validation (10%), and test (20%) sets, following ChestX-ray14.

2. Results

2.1. Binary Classification: CheXNet vs Radiologists

CheXNet achieves an F1 score of 0.435 (95% CI 0.387, 0.481), higher than the radiologist average of 0.387 (95% CI 0.330, 0.442).

2.2. 14-Pathology Classification: CheXNet vs SOTA Approaches

CheXNet achieves state of the art results on all 14 pathology classes.

2.3. Visualization

• Class Activation Mappings (CAM) is used for visualization.
• The above Figure 2 shows several examples of CAMs on the pneumonia detection task as well as the 14-class pathology classification task.