[Paper Review] Accurate Lung Segmentation via Network-Wise Training of Convolutional Networks

Sangheum Hwang, Sunggyun Park|arXiv (Cornell University)|Aug 2, 2017

COVID-19 diagnosis using AI3 references3 citations

TL;DR

This paper proposes a deep convolutional neural network with atrous convolutions and a novel network-wise training strategy for accurate lung segmentation in chest X-rays. By iteratively refining predictions using outputs from prior stages, the method reduces false positives and negatives and produces smoother boundaries, achieving state-of-the-art performance on the JSRT dataset with fewer parameters and improved generalization when combined with data augmentation.

ABSTRACT

We introduce an accurate lung segmentation model for chest radiographs based on deep convolutional neural networks. Our model is based on atrous convolutional layers to increase the field-of-view of filters efficiently. To improve segmentation performances further, we also propose a multi-stage training strategy, network-wise training, which the current stage network is fed with both input images and the outputs from pre-stage network. It is shown that this strategy has an ability to reduce falsely predicted labels and produce smooth boundaries of lung fields. We evaluate the proposed model on a common benchmark dataset, JSRT, and achieve the state-of-the-art segmentation performances with much fewer model parameters.

Motivation & Objective

• To develop a highly accurate, parameter-efficient lung segmentation model for chest radiographs using deep learning.
• To address the challenge of high variability in lung shape and size across X-ray images.
• To improve segmentation accuracy by reducing false positives and false negatives and producing smoother lung boundaries.
• To investigate the effectiveness of a multi-stage training strategy—network-wise training—where each stage uses both input images and outputs from previous stages.
• To evaluate generalization performance on a distribution-shifted dataset (MC) and assess robustness to domain shift.

Proposed method

• Uses a deep-and-thin residual network architecture with atrous convolutions (rate r=3) to expand receptive fields without increasing parameters.
• Applies batch normalization after every convolutional layer to stabilize training.
• Introduces network-wise training: each subsequent stage network is trained on both original images and the segmentation outputs from the previous stage.
• Employs a multi-scale loss function combining cross-entropy and boundary loss to improve contour accuracy.
• Applies data augmentation by randomly adjusting brightness and contrast to improve robustness to pixel value variations.
• Uses a U-Net-like decoder with transposed convolutions and skip connections to restore spatial resolution.

Experimental results

Research questions

• RQ1 Can network-wise training, where each stage is trained on both input images and prior stage outputs, improve lung segmentation accuracy and boundary smoothness?
• RQ2 Does the use of atrous convolutions in a lightweight network architecture achieve competitive performance with fewer parameters?
• RQ3 How does the model perform on a distribution-shifted dataset (MC) compared to the training dataset (JSRT), and what does this imply for domain shift robustness?
• RQ4 To what extent does data augmentation improve generalization and reduce variance in segmentation performance on out-of-distribution data?
• RQ5 Can the proposed method achieve state-of-the-art performance across multiple metrics (Jaccard, Dice, ACD, ASD) while maintaining model efficiency?

Key findings

• The proposed model achieves state-of-the-art performance on the JSRT dataset, with a Jaccard similarity coefficient (JSC) of 0.961 ± 0.015 and Dice coefficient (DC) of 0.980 ± 0.008 when using data augmentation at stage 3.
• Network-wise training significantly reduces false positives and negatives, with boundary smoothness improving from stage 1 to stage 3, as shown in visual comparisons.
• Performance saturates at stage 3, indicating diminishing returns beyond three stages, so stage 3 is used for final evaluation.
• On the MC dataset, the model shows lower mean performance (JSC: 0.931 ± 0.049) and higher variance (standard deviation 0.049), indicating instability due to domain shift.
• The model achieves a mean surface distance (ASD) of 0.675 ± 0.122 mm on JSRT with data augmentation, outperforming prior methods.
• The method maintains a small model size due to its deep-and-thin architecture, achieving high accuracy with fewer parameters than competing models.