Source-linked AI summary
ChestNet: A Deep Neural Network for Classification of Thoracic Diseases on Chest Radiography
Hongyu Wang, Yong Xia
TL;DR
Chest radiograph diagnosis is challenging, and abnormality locations are often unavailable with disease labels. ChestNet combines classification and attention branches, and outperformed methods using no extra training data on the official patient-wise ChestX-ray14 split.
Problem
Chest radiograph diagnosis is challenging because abnormality labels often identify diseases without specifying their locations, making the task weakly supervised.
Source: §1 · pp. 1–2Method
ChestNet combines a ResNet-152 classification branch with an attention branch that relates class labels to pathological-abnormality locations.
Source: §1 · p. 2, §3.1 · p. 3, §3.2 · p. 4Results
0.781 average per-class AUC; ChestNet outperformed methods using no extra training data on the official patient-wise ChestX-ray14 split.
Source: §6 · p. 7Takeaways & Limitations
Inference took less than 0.1 second per radiograph on the reported hardware, which the authors say could suit routine clinical workflow.
Source: §5 · p. 7
Abstract
from arXiv · showhide
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.
1 Introduction
Chest radiograph diagnosis is demanding for human readers, while limited abnormality-location labels make automated diagnosis a weakly supervised problem. ChestNet addresses this with classification and attention branches that use class-label and abnormality-location correlations.
1 Introduction
- Manual slice-by-slice reading is time-consuming, expensive, prone to operator bias, and difficult even for radiologists.The paper also notes the high skill and concentration required.Source: §1 · pp. 1–2
- Abnormality localization is weakly supervised because images provide disease labels but generally lack abnormality locations, and few public datasets include masks.Source: §1 · p. 2
- Attention-based models can learn localization and recognition of multiple objects from class labels, motivating attention for this task.Source: §1 · p. 2
- ChestNet combines a classification branch with an attention branch that relates class labels to abnormality locations.The classification branch provides feature extraction and classification without handcrafted features.Source: §1 · p. 2
2 Dataset
ChestX-ray14 is a patient-wise split dataset of frontal chest radiographs with 14 disease labels mined from radiology reports.
2 Dataset
- 112,120 frontal-view images from 30,805 subjects make up the hospital-scale ChestX-ray14 dataset.Images were rescaled to 1024 × 1024 pixels.Source: §2 · p. 3
- The dataset contains 14 disease labels mined from associated radiological reports using natural language processing.Images may have one or multiple labels.Source: §2 · p. 3
- The official split assigns 80% of subjects to training and 20% to testing, keeping each patient's studies in only one subset.Source: §2 · p. 3
3 Method
ChestNet combines a ResNet-152 classification branch with an attention branch for abnormality detection. It is trained in stages and then fine-tuned end-to-end, with the two branch outputs averaged for the full model.
3 Method
- ChestNet consists of a classification branch and an attention branch.Source: §3 · p. 3, Figure 1 · p. 3
3.1 Classification Branch for Label Prediction
- The classification branch adapts ImageNet-pretrained ResNet-152 to produce 14 sigmoid disease-label predictions.The final fully connected layer is replaced by 14 neurons.Source: §3.1 · p. 3
3.2 Attention Branch for Abnormality Detection
- The attention branch uses the penultimate ResNet-152 module's features and a normalized Grad-CAM localization map to estimate class confidence.The normalized map is fed into the later convolutional layers of the attention branch.Source: §3.2 · p. 4
3.3 Training and Testing
- Radiographs were resized to 224 × 224 without augmentation; training used a patient-wise split with 10% of training subjects for validation, then classification, fixed-classifier attention, and end-to-end fine-tuning.For each test image, any disease prediction above 0.5 was positive.Source: §3.3 · pp. 4–5
4 Results
ChestNet achieved the most accurate per-disease diagnosis among compared methods trained only on ChestX-ray14, while ranking second in average per-class AUC overall. The section also presents example diagnoses and ROC curves for all 14 diseases.
4 Results
- The two test examples were diagnosed as atelectasis with effusion, and effusion with infiltration, respectively.Source: §4 · p. 5, Figure 2 · p. 5
- ROC curves are shown for ChestNet's diagnosis of each of the 14 thorax diseases.Source: §4 · p. 5, Figure 3 · p. 5
- ChestNet achieved the second-highest average per-class AUC overall and the most accurate diagnosis of all 14 diseases among methods trained only on ChestX-ray14.The [15] method used more than 180,000 PLCO images as extra training data.Source: Table 1 · p. 6, §4 · p. 6
5 Discussion
The attention branch is associated with a higher average per-class AUC than ResNet-152, while ChestNet’s training is computationally complex but inference is fast.
Contribution of the Attention Branch
- Average per-class AUC rose from 0.7401 with ResNet-152 to 0.7810 with ChestNet.The comparison evaluates the contribution of the attention branch under the same experimental settings.Source: §5 · p. 6, Table 2 · p. 7
- The attention branch concentrates adaptively on pathological regions and is described as critical to ChestNet’s performance.Removing it would leave the ResNet-152 classification branch.Source: §5 · p. 6
Computational Complexity
- Training ChestNet took about 20 hours, while diagnosing one radiograph took less than 0.1 second on the reported hardware.The authors suggest this separation of offline training and rapid online testing could fit a routine clinical workflow.Source: §5 · p. 7
6 Conclusion
ChestNet combines classification and attention branches to diagnose 14 thorax diseases on chest radiographs. On ChestX-ray14, it achieved a higher average per-class AUC than other methods using no extra training data; future work will incorporate label correlations.
6 Conclusion
- ChestNet combines classification and attention branches to diagnose 14 thorax diseases on chest radiographs.It was evaluated against three state-of-the-art deep learning models on ChestX-ray14 using the official patient-wise split.Source: §6 · p. 7
- ChestNet achieved an average per-class AUC of 0.781, higher than other methods using no extra training data.Source: §6 · p. 7
- Future work will incorporate correlations among disease image labels into computer-aided diagnosis.Source: §6 · p. 7