Source-linked AI summary
Generalization of Deep Neural Networks for Chest Pathology Classification in X-Rays Using Generative Adversarial Networks
Hojjat Salehinejad, Shahrokh Valaee, Tim Dowdell, Errol Colak, Joseph Barfett
TL;DR
Medical datasets often underrepresent rare conditions and overrepresent common ones, making balanced pathology classification difficult. The paper generates synthetic chest X-rays with a DCGAN and combines them with real images to train a DCNN across five classes. The augmented dataset improves classification performance, with a reported mean accuracy of 92.10% and better generalization than training on real data alone.
Problem
Medical datasets are often imbalanced, with common conditions overrepresented and rare conditions underrepresented, limiting balanced pathology classification.
Method
A DCGAN generates synthetic chest X-rays, which are combined with real images to balance and expand training data for a DCNN classifying five pathological classes.
Results
92.10% mean classification accuracy was achieved using DCGAN-augmented data, outperforming the original dataset across all classes.
Takeaways & Limitations
Synthesized images increase training-set diversity and balance, improving reported generalization performance for chest pathology classification of unseen data.
Abstract
from arXiv · showhide
Medical datasets are often highly imbalanced with over-representation of common medical problems and a paucity of data from rare conditions. We propose simulation of pathology in images to overcome the above limitations. Using chest X-rays as a model medical image, we implement a generative adversarial network (GAN) to create artificial images based upon a modest sized labeled dataset. We employ a combination of real and artificial images to train a deep convolutional neural network (DCNN) to detect pathology across five classes of chest X-rays. Furthermore, we demonstrate that augmenting the original imbalanced dataset with GAN generated images improves performance of chest pathology classification using the proposed DCNN in comparison to the same DCNN trained with the original dataset alone. This improved performance is largely attributed to balancing of the dataset using GAN generated images, where image classes that are lacking in example images are preferentially augmented.
1. INTRODUCTION
Medical datasets are constrained by privacy, limited access, and class imbalance, especially for rare conditions. The paper proposes controlled pathology simulation with GANs to augment and balance chest X-ray data for DCNN training.
- Privacy concerns and restricted research access make medical data intrinsically limited.
- Common hospital conditions are overrepresented, while rare syndromes such as Birt-Hogg-Dube syndrome have limited clinical data.
- Class imbalance makes training neural networks with similarly high accuracy across classes technically challenging.
- Data augmentation can partially address uneven image availability by applying transformations to existing images.
- GANs can generate artificial data that is described as indiscernible from real counterparts, motivating their use for medical-image augmentation.
- The proposed DCGAN generates chest X-rays mimicking common pathologies to balance and augment labeled data across five pathological classes.
2. PROPOSED METHOD
The proposed method combines a DCGAN that generates synthetic chest X-rays with a DCNN that classifies abnormalities using real and generated images. The pipeline uses adversarial image generation, dataset concatenation, convolutional feature extraction, pooling, and five-class softmax classification.
- Generating Chest X-Rays: The Generator and Discriminator compete adversarially: the Discriminator maximizes the loss, while the Generator minimizes it to produce images indistinguishable from real X-rays.Each network is trained in a feed-forward fashion while the other is trained, without back propagation through the inactive component.
- Classification of Chest X-Rays: Real and synthesized chest X-rays are shuffled and concatenated before being fed to the proposed DCNN for pathology detection and classification.This combines the DCGAN generation stage with the downstream chest-X-ray classification stage.
- Classification of Chest X-Rays: The proposed DCNN is broadly AlexNet-like, using convolutional layers for feature extraction and fully connected layers for classification, with modified kernels, feature-map sizes, and convolution layers.For 256 × 256 inputs, the architecture applies 5 × 5 convolution kernels, ReLU activations, and max-pooling before the classifier.
- Classification of Chest X-Rays: Max-pooled features are reshaped into a vector and passed to a one-hidden-layer MLP whose softmax output represents five chest X-ray classes.The classes are Pneumothorax, Pulmonary Edema, Pleural Effusion, Normal, and Cardiomegaly.
3. EXPERIMENTS
The experiments evaluate DCGAN-generated chest X-rays as augmentation for a DCNN, using real imbalanced, real balanced, and synthesized-image-balanced datasets. Radiologist review, dataset composition, classification accuracy, and validation behavior assess image plausibility and classification performance.
- Dataset and evaluation: The hospital dataset contained 15,781 Normal, 17,098 Cardiomegaly, 14,510 Pleural Effusion, 5,018 Pulmonary Edema, and 4,013 Pneumothorax exams.Images were anonymized and down-sampled to 256×256 pixels; 1,000 equal-class images were reserved for validation and another 1,000 for testing.
- Dataset and evaluation: Generated images were reviewed by a board-certified radiologist and combined with real images in varied proportions before DCNN assessment on test data.The evaluation examined both qualitative class-appropriate features and quantitative classification performance.
- Qualitative evaluation: The artificial and real chest X-rays showed similar characteristics, although the synthetic images had comparatively low resolution.Examples displayed class-related features including pulmonary edema attenuation, pneumothorax pleural lines, cardiomegaly, and bilateral pleural effusions.
- Dataset balancing: Each class in the balanced dataset contained 30,196 images, twice the maximum available real-class count, with the additional samples supplied by DCGAN synthesis.The balanced total combined real and synthesized images, preferentially expanding underrepresented classes.
- Classification results: 92.10% mean classification accuracy was achieved with DCGAN augmentation, more than 20% above the same DCNN trained only on the original data.The greatest improvement occurred for pneumothorax, the class with the fewest native images.
- Training behavior: The DS3-trained DCNN reached higher validation accuracy than DS1 and DS2, nearly converged after 90 iterations, while DS1 and DS2 overfitted after 22 and 31 iterations.The reported comparison links the added generated-image diversity with improved generalization and reduced over-fitting.
4. CONCLUSION
The conclusion reports that DCGAN-generated images can expand and balance real chest X-ray datasets for DCNN training. The best results came from combining real and artificial data, with reported gains concentrated in the most data-sparse classes.
- Conclusion: DCGAN-generated artificial data provided greater training quantity and balanced the dataset, improving classification performance substantially in the most anemic classes.The conclusion states that the strongest results used a combination of real and artificial data.
- Conclusion: The reported results suggest that synthesized-image augmentation increases training-set diversity and improves DCNN generalization on unseen chest X-ray data.