Source-linked AI summary
Attention U-Net Based Adversarial Architectures for Chest X-ray Lung Segmentation
Gusztáv Gaál, Balázs Maga, András Lukács
TL;DR
Lung segmentation is an arduous step in the diagnostic pipeline, with pixel-wise segmentation methods and adversarial approaches already explored. This work combines an Attention U-Net with an adversarial critic and reports a lung segmentation Dice score of 0.962, while noting cross-dataset performance and subtle shape improvements from the adversarial scheme.
Problem
Lung segmentation is an arduous diagnostic-pipeline task, and semantic segmentation requires assigning a predefined class to each pixel.
Source: §1 · p. 1Method
The approach uses an Attention U-Net with an adversarial critic and replaces multiclass cross-entropy with Focal Tversky Loss in the segmentation loss.
Source: §2.1 · p. 2, §2.3 · p. 4Results
0.962 dice score was reached for lung segmentation using the Attention U-Net infrastructure.
Source: §4 · p. 6Takeaways & Limitations
With preprocessing, the network performed well when test and validation sets came from different datasets, while the adversarial scheme provided subtle shape improvements.
Source: §4 · p. 6The critic architecture was not experimentally varied; further tests are proposed to better understand the adversarial scheme's role.
Source: §5 · p. 6
Abstract
from arXiv · showhide
Chest X-ray is the most common test among medical imaging modalities. It is applied for detection and differentiation of, among others, lung cancer, tuberculosis, and pneumonia, the last with importance due to the COVID-19 disease. Integrating computer-aided detection methods into the radiologist diagnostic pipeline, greatly reduces the doctors' workload, increasing reliability and quantitative analysis. Here we present a novel deep learning approach for lung segmentation, a basic, but arduous task in the diagnostic pipeline. Our method uses state-of-the-art fully convolutional neural networks in conjunction with an adversarial critic model. It generalized well to CXR images of unseen datasets with different patient profiles, achieving a final DSC of 97.5% on the JSRT dataset.
1 Introduction
Chest X-rays are widely used, and lung segmentation supports nodule identification and other diagnostic tasks. The introduction frames segmentation as pixel-wise classification and situates the work among neural-network and adversarial approaches.
1 Introduction
- Chest X-ray is the most commonly requested imaging test by general practitioners, and lung segmentation is essential for identifying lung nodules and calculating the cardiothoracic ratio.The cardiothoracic ratio is described as the primary indicator of cardiomegaly.Source: §1 · p. 1
- Semantic segmentation assigns a predefined class to each image pixel, and fully convolutional networks attain state-of-the-art performance on this task.Source: §1 · p. 1
- Prior chest X-ray work explored adversarial training with fully convolutional networks and Mask R-CNN for instance segmentation.The cited work reports state-of-the-art results for Mask R-CNN-based chest X-ray instance segmentation.Source: §1 · p. 1
2 Deep Learning Approach
This section combines Attention U-Net’s attention gates with adversarial shape correction, and uses Focal Tversky Loss in place of cross-entropy for segmentation training.
2.1 Network Architecture
- Attention U-Net produces input-resolution masks and uses attention gates to focus on relevant local features, avoiding a separate ROI-selection network.The authors also report that suppressing less relevant features reduces false positives.Source: §2.1 · p. 2
- The adversarial variant replaces SCAN’s FCN with Attention U-Net while leaving the critic design unchanged.Further experiments with critic architecture are left for future work.Source: §2.1 · p. 2
2.2 Tversky Loss
- Tversky similarity weights false negatives and false positives separately, addressing Dice Loss’s equal penalties and its associated high-precision, low-recall tendency.The passage notes that false negatives may need greater weight than false positives when regions of interest are small.Source: §2.2 · p. 3
- Focal Tversky Loss targets the difficulty of small regions contributing little to Dice Loss, with γ in [1, 3] emphasizing misclassified pixels with small Tversky index.The passage says the loss decreases significantly for such misclassified pixels.Source: §2.2 · p. 4
2.3 Training
- Training alternates between a critic that distinguishes ground-truth from predicted masks and a segmentation network trained with Focal Tversky Loss plus a weighted critic term.The critic term uses λ to balance pixel-wise and adversarial losses.Source: §2.3 · p. 4
- The training schedule pretrains the generator for 50 epochs, then trains adversarially for 50 epochs with one critic step per five generator steps.The authors report that their network trains much faster than the schedule borrowed from.Source: §2.3 · p. 4
3 Datasets
The experiments use three public chest X-ray datasets with expert-reviewed segmentation masks. Preprocessing applies CLAHE to improve local contrast, then resizes images to 512×512 and maps them to [−1, 1].
3 Datasets
- JSRT has 247 images (154 with nodules), Montgomery 138 (80 healthy and 58 with tuberculosis), and Shenzhen 662 (326 healthy and 336 with tuberculosis).JSRT includes lung and heart masks; Montgomery and Shenzhen provide lung masks.Source: §3 · p. 5
- All three datasets are public and have organ segmentation masks reviewed by expert radiologists.Source: §3 · p. 5
3.1 Preprocessing Data
- CLAHE clips histogram values before redistribution to enhance local contrast while counteracting noise amplification in near-constant regions.The text says the preprocessing also helps neural networks; Figure 3 shows example images before and after CLAHE.Source: §3.1 · p. 5, Figure 3 · p. 5
- Images were resized to 512×512 and mapped to [−1, 1] before network input.Source: §3.1 · p. 5
4 Experiments and Results
Attention U-Net achieved a lung-segmentation Dice score of 0.962 and performed well when validation and test sets came from different datasets. Adversarial training increased Dice scores, but its gains were modest relative to those reported for FCN and SCAN.
- Training used 40 epochs, batches of eight, SGD, and Focal Tversky Loss; Adam failed to converge in many cases.Source: §4 · p. 5
- Flipping, rotating, shearing, and changing image brightness did not help and resulted in slower convergence.Source: §4 · p. 6
- A Dice score of 0.962 was achieved for lung segmentation, with good performance when validation and test sets came from different datasets.The paper notes that this cross-dataset performance matters because X-rays can differ substantially between machines.Source: §4 · p. 6
- Adversarial training increased Dice scores, but the improvement was less pronounced than for FCN and SCAN.The authors attribute the subtler gains to preprocessing that prevents Attention U-Net's vanilla masks from showing the abnormal shapes seen with FCN, leaving subtle shape improvements.Source: §4 · p. 6
5 Future Work
The authors leave critic-network architecture experiments for future work, despite finding the architecture from satisfactory.
- The critic architecture from was not varied; the authors recommend further tests to better understand the adversarial scheme's role.Source: §5 · p. 6