[Paper Review] ChestNet: A Deep Neural Network for Classification of Thoracic Diseases on Chest Radiography

Hongyu Wang, Yong Xia|arXiv (Cornell University)|Jul 9, 2018

COVID-19 diagnosis using AIMedicine15 references101 citations

TL;DR

ChestNet combines a classification branch with an attention branch to diagnose 14 thoracic diseases from chest X-rays, outperforming state-of-the-art models on ChestX-ray14 without using extra training data.

ABSTRACT

Computer-aided techniques may lead to more accurate and more acces-sible diagnosis of thorax diseases on chest radiography. Despite the success of deep learning-based solutions, this task remains a major challenge in smart healthcare, since it is intrinsically a weakly supervised learning problem. In this paper, we incorporate the attention mechanism into a deep convolutional neural network, and thus propose the ChestNet model to address effective diagnosis of thorax diseases on chest radiography. This model consists of two branches: a classification branch serves as a uniform feature extraction-classification network to free users from troublesome handcrafted feature extraction, and an attention branch exploits the correlation between class labels and the locations of patholog-ical abnormalities and allows the model to concentrate adaptively on the patholog-ically abnormal regions. We evaluated our model against three state-of-the-art deep learning models on the Chest X-ray 14 dataset using the official patient-wise split. The results indicate that our model outperforms other methods, which use no extra training data, in diagnosing 14 thorax diseases on chest radiography.

Motivation & Objective

• Motivate accurate and accessible diagnosis of thorax diseases from chest radiographs using deep learning.
• Address the weakly supervised nature of thoracic disease localization and diagnosis.
• Eliminate the need for handcrafted feature extraction by using a unified feature extraction-classification network.

Proposed method

• Introduce a two-branch architecture: a classification branch for uniform feature extraction and classification, and an attention branch that links class labels to abnormal locations.
• Incorporate an attention mechanism to exploit the correlation between labels and pathologically abnormal regions, allowing adaptive focus on diseased areas.
• Leverage the ChestX-ray14 dataset with the official patient-wise split to evaluate performance.
• Avoid using any extra training data beyond the dataset to compare with state-of-the-art models.

Experimental results

Research questions

• RQ1 Can attention mechanisms improve localization and diagnostic accuracy for thoracic diseases in chest radiographs?
• RQ2 Does ChestNet achieve superior diagnosis performance on ChestX-ray14 without relying on additional training data?
• RQ3 Is a two-branch architecture with an attention component more effective than single-branch approaches with handcrafted features?
• RQ4 How does the model perform across the 14 thoracic disease categories in the ChestX-ray14 dataset?

Key findings

• ChestNet outperforms three state-of-the-art deep learning models on ChestX-ray14 using the official patient-wise split.
• The model achieves better disease diagnosis performance without requiring extra training data beyond ChestX-ray14.
• The attention branch enables the network to focus on pathologically abnormal regions by leveraging the correlation between labels and lesion locations.
• The proposed framework reduces reliance on handcrafted features by providing a uniform feature extraction-classification pipeline.