PaperlayerView exact revision on arXiv

Source-linked AI summary

Abnormality Detection and Localization in Chest X-Rays using Deep Convolutional Neural Networks

Mohammad Tariqul Islam, Md Abdul Aowal, Ahmed Tahseen Minhaz, Khalid Ashraf

arXiv:1705.09850v3cs.CV

TL;DR

Chest X-ray abnormality detection lacks standardized public benchmarks, limiting comparison across methods. This paper evaluates deep convolutional networks on public datasets and finds that architecture choice, shallow features, and model ensembling matter. It reports leading detection results, including a 17-percentage-point cardiomegaly gain over rule-based methods and highest tuberculosis accuracy on the Shenzhen dataset. Localization also identifies abnormal regions for spatially spread-out diseases, while the paper notes that its evaluation should be extended to newly announced datasets.

  • Problem

    The central problem is limited comparison of chest X-ray detection methods because standard publicly available datasets and benchmarks are lacking.

  • Method

    The study evaluates deep convolutional network architectures, feature depths, ensembles, and localization methods on publicly available chest X-ray datasets.

  • Results

    The study finds that shallow or earlier-layer features and deep-network ensembles improve classification, reports a 17-percentage-point cardiomegaly gain over rule-based methods, and achieves highest tuberculosis accuracy on the Shenzhen dataset.

  • Takeaways & Limitations

    Deep-learning classification and localization can identify abnormalities in chest X-rays, including spatially spread-out diseases, within the evaluated datasets.

  • Takeaways & Limitations

    The evaluation is based on publicly available datasets and frontal chest X-rays, and the authors identify applying the techniques to a newly announced dataset as future work.

Abstract

from arXiv · show

Chest X-Rays (CXRs) are widely used for diagnosing abnormalities in the heart and lung area. Automatically detecting these abnormalities with high accuracy could greatly enhance real world diagnosis processes. Lack of standard publicly available dataset and benchmark studies, however, makes it difficult to compare various detection methods. In order to overcome these difficulties, we have used a publicly available Indiana CXR, JSRT and Shenzhen dataset and studied the performance of known deep convolutional network (DCN) architectures on different abnormalities. We find that the same DCN architecture doesn't perform well across all abnormalities. Shallow features or earlier layers consistently provide higher detection accuracy compared to deep features. We have also found ensemble models to improve classification significantly compared to single model. Combining these insight, we report the highest accuracy on chest X-Ray abnormality detection on these datasets. We find that for cardiomegaly detection, the deep learning method improves the accuracy by a staggering 17 percentage point compared to rule based methods. We applied the techniques to the problem of tuberculosis detection on a different dataset and achieved the highest accuracy. Our localization experiments using these trained classifiers show that for spatially spread out abnormalities like cardiomegaly and pulmonary edema, the network can localize the abnormalities successfully most of the time. One remarkable result of the cardiomegaly localization is that the heart and its surrounding region is most responsible for cardiomegaly detection, in contrast to the rule based models where the ratio of heart and lung area is used as the measure. We believe that through deep learning based classification and localization, we will discover many more interesting features in medical image diagnosis that are not considered traditionally.

1. Introduction

The paper addresses limited comparability in chest X-ray abnormality detection by evaluating deep convolutional networks on public datasets. It reports architecture- and feature-dependent performance, ensemble gains, and strong results for cardiomegaly and tuberculosis detection.

  • Motivation and contributions: Public datasets and benchmark results are used to address limited comparability among chest X-ray abnormality detection methods.The study uses publicly available Indiana, JSRT, and Shenzhen datasets.
  • Key results: 17 percentage points improvement in cardiomegaly accuracy is achieved over rule-based methods using an ensemble of deep convolutional networks.
  • Key results: Multiple random train/test splits produce robust accuracy results when training examples are limited.
  • Key results: Shallow features or earlier network layers perform better than deep features for classification accuracy.
  • Key results: Ensembles of deep convolutional networks outperform single models, whereas combining them with rule-based models degrades accuracy.
  • Localization and transfer: Sensitivity-based localization correctly identifies regions for spatially spread-out diseases, and the study reports highest accuracy for tuberculosis detection on the Shenzhen dataset.

2. Related Works

Prior chest X-ray CAD studies used handcrafted features, segmentation, registration, and rule-based measures, while some relied on small private datasets. Related work also includes deep learning applications in other medical imaging tasks.

  • Chest X-ray CAD: Earlier CAD work used local binary pattern features on segmented chest X-rays to classify normal and pathological images.The cited study used a private dataset containing 48 images.
  • Chest X-ray CAD: Rule-based cardiomegaly detection used heart and lung registration followed by cardiothoracic ratio and cardiothoracic area ratio measurements.
  • Deep learning in medical imaging: Related medical-imaging studies applied deep learning to localization and classification tasks including cancer cells, diabetic retinopathy, and skin cancer.

3. Experiments

The experiments use three public CXR datasets to evaluate DCN-based abnormality classification and occlusion-sensitivity localization. Detection is assessed with accuracy, AUC, sensitivity, and specificity, while localization maps identify regions important for classification.

  • Datasets: The study uses Indiana, JSRT, and Shenzhen datasets containing disease annotations, lung nodules, or normal-versus-tuberculosis labels.Indiana includes cardiomegaly and other abnormalities; JSRT provides gold-standard masks for nodule evaluation; Shenzhen contains 662 frontal CXRs.
  • Dataset examples: The Indiana examples contrast a normal CXR with cardiomegaly whose abnormal heart size and shape are visually distinguishable.The cardiomegaly example shows a heart that is quite large compared with the normal example.
  • Classification models: The experiments evaluate DCN architectures on binary classification of cardiomegaly and pulmonary atelectasis against normal chest X-rays.The explored models include AlexNet, VGG-Net, and ResNet, which differ in convolutional depth.
  • Evaluation metrics: Detection quality is measured using accuracy, AUC, sensitivity, and specificity, with a 0.50 threshold for reported accuracy, sensitivity, and specificity unless stated otherwise.ROC curves vary the classifier threshold from 0 to 1; sensitivity is true positive rate, while specificity measures correct identification of normal samples.
  • Localization method: The classification ensemble combines different DCN models, while localization exploits overlapping occlusions and occlusion sensitivity.Sliding an occluded square patch across the CXR produces a pathology probability map; low-probability regions indicate likely pathology locations.

4.1. Classification

Classification performance depends on architecture, feature depth, regularization, training-set size, and model combination. Earlier-layer features and ensembles are consistently valuable, while the best architecture varies by abnormality.

  • Regularization: Dropout improves shallower networks but reduces deep-model performance by about 4 percentage points.With dropout, VGG-16 achieves the highest accuracy and AlexNet the highest AUC, while ResNet-101 and VGG-19 decline.
  • Feature depth: Earlier-layer features improve accuracy by 2 to 4 percentage points compared with later-layer features.For ResNet-152 cardiomegaly detection, pool5 underperforms earlier layers, with residual-connection and ReLU features performing considerably better.
  • Feature fusion: Fusing rule-based features with DCN features degrades classification performance and is therefore not reported.The fusion used CTR, 2D-CTR, and CTAR alongside DCN features.
  • Architecture choice: Different DCNs maximize different classification figures of merit, so no single architecture is uniformly best across abnormalities.For pulmonary edema, the results indicate that deeper models are not consistently better; ResNet-50 performs consistently better.
  • Training-set size: Accuracy and AUC stabilize at about 200 training images as the mean rises and result variability falls.With fewer than 50 examples per category, the mean is lower and deviation is higher.
  • Ensembling: Ensembles of DCN models improve robustness and separate pathology from normal classes more effectively than individual models.At one operating point, the ensemble reaches 98% sensitivity with 82% specificity; at another, 98% specificity with 86% sensitivity.

4.2. Localization

The paper uses sensitivity-based patch occlusion to localize regions contributing to abnormality predictions. Localization is consistent for cardiomegaly and works best within lung regions for pulmonary edema.

  • Method: Sensitivity-based localization provides correct localization for spatially spread-out diseases.The method identifies image regions whose occlusion most changes classification confidence.
  • Cardiomegaly: For cardiomegaly, the highest-sensitivity region is concentrated around the heart in most examples.The experiment used 50 cardiomegaly and 50 normal images, showing consistent localization for most examples.
  • Cardiomegaly: Cardiomegaly localization indicates that heart shape and surrounding regions contribute more than lung-relative-size features.This contrasts with rule-based CTR and CTAR measures and is identified as requiring further exploration.
  • Pulmonary edema: Pulmonary-edema localization is obtained best when lung regions of interest are used to compute the probability map.The localized area corresponds to excess fluid in the lungs, without an associated anatomical shape change.

4.3. Comparison between Rule based and DCN based cardiomegaly detection

DCN-based cardiomegaly detection outperforms the reproduced rule-based baseline, with the linear ensemble achieving the strongest reported accuracy and AUC.

  • Performance comparison: 93% accuracy and 0.9728 AUC are achieved by the linear ensemble of six DCN models.VGG-19 is the strongest independent model, reaching 92% accuracy and 0.9408 AUC.
  • Performance comparison: The ensemble improves accuracy by 17 percentage points over the reproduced rule-based method.The rule-based baseline achieves 75.6% accuracy, compared with 93% for the ensemble.

4.4. Tuberculosis Detection

The authors evaluate their DCN pipeline on the Shenzhen tuberculosis dataset alongside prior methods. The ensemble obtains the highest reported accuracy and AUC in that comparison.

  • Evaluation: The Shenzhen dataset is used to assess tuberculosis detection with DCN models and their ensemble.A detailed tuberculosis study is deferred to future work.
  • Results: 90% accuracy and 0.94 AUC are obtained by the ensemble method at a classifier threshold of 0.74.At threshold 0.50, accuracy is 88%; the ensemble provides the highest accuracy and AUC in the comparison.
  • Results: VGG-16 achieves the highest sensitivity, while AlexNet achieves the highest specificity among the reported DCN models.Sensitivity and specificity for several prior methods are unavailable because those papers did not report them.

5. Conclusion

The paper addresses limited comparability in chest X-ray abnormality detection by evaluating deep convolutional methods on public datasets. It finds that shallow features, repeated random splits, and ensembles improve detection, while localization succeeds mainly for spatially spread abnormalities.

  • Public datasets and repeated random train-test splits address limited comparability and stabilize accuracy estimates when training examples are scarce.
  • Shallow or earlier-layer features consistently outperform deeper features for chest X-ray abnormality classification.
  • Ensembles significantly improve classification over single DCNN models, whereas combining DCNN and rule-based models degrades accuracy.
  • 17 percentage points: deep learning improves cardiomegaly accuracy over rule-based methods.
  • Localization usually succeeds for spatially spread abnormalities such as cardiomegaly and pulmonary edema but fails for pointed features such as nodules or fractures.
  • The paper identifies a boundary for future evaluation: a newly announced dataset was not yet available for applying the proposed techniques.

7. Supplementary Materials

The supplementary methods describe model-based lung-heart segmentation, classification evaluation, and performance analysis across 20 chest X-ray abnormalities. They combine visual similarity, dense correspondence, and transformed anatomical boundaries to estimate segmentation.

  • The supplementary workflow covers CXR segmentation, rule-based modeling, classification of 20 abnormalities, and additional localization results.
  • Model-based segmentation registers radiologist-marked lung and heart boundaries from existing CXRs to a patient X-ray.
  • Radon transforms and Bhattacharyya distance identify five visually similar JSRT samples for each test CXR.
  • SIFT-flow matches densely sampled local features to transform model-X-ray boundaries into approximate lung-heart segmentation for the test image.
  • Rule-based features and SVM classify cardiomegaly versus normal CXRs, while ResNet-152 evaluation reports accuracy, sensitivity, and specificity.
  • ResNet-152 performance is analyzed for 20 abnormalities using balanced normal-image sampling and a 70:30 train-test split.

8. Additional Examples of Localization

Additional localization examples show disease-specific concentration for cardiomegaly and pulmonary edema, contrasted with dispersed patterns in normal images. Misclassified cases demonstrate that useful localization can coexist with incorrect classification.

  • Cardiomegaly cases show fine localization around the heart, whereas normal CXRs show dispersed low-probability regions.
  • Cardiomegaly heat maps remain sensitive under occlusion and indicate reliance on local rather than broadly distributed features.
  • A cardiomegaly-positive CXR misclassified as normal still localizes the heart region accurately.
  • A normal CXR misclassified as cardiomegaly shows stronger heart localization than typical normal images but retains a spread-out pattern.
  • Pulmonary edema examples generally localize in the lung region, unlike normal images with random dense localization.
  • Pulmonary edema localization can vary: one normal-like example localizes in the lungs, while another resembles normal-CXR patterns.
Loading 1705.09850v3…