PaperlayerView exact revision on arXiv

Source-linked AI summary

Pneumonia Detection on chest X-ray images Using Ensemble of Deep Convolutional Neural Networks

Alhassan Mabrouk, Rebeca P. Díaz Redondo, Abdelghani Dahou, Mohamed Abd Elaziz, Mohammed Kayed

arXiv:2312.07965v1eess.IVcs.CVcs.LG

TL;DR

Pneumonia diagnosis from chest X-rays is difficult, making automated classification important. The paper proposes an ensemble combining three pretrained models and reports better test-set accuracy with lower loss.

  • Problem

    Automated pneumonia classification from chest X-rays is important because early detection is critical for this serious lung infection.

  • Method

    The proposed Ensemble Learning method combines MobileNetV2, DenseNet169, and Vision Transformer models trained for pneumonia detection in chest X-ray images.

  • Results

    The proposed method achieved better test-set accuracy and lower loss than the individual models.

  • Takeaways & Limitations

    Combining three CNN-based models with a Vision Transformer lightly improves state-of-the-art performance using relatively few layers and features.

Abstract

from arXiv · show

Pneumonia is a life-threatening lung infection resulting from several different viral infections. Identifying and treating pneumonia on chest X-ray images can be difficult due to its similarity to other pulmonary diseases. Thus, the existing methods for predicting pneumonia cannot attain substantial levels of accuracy. Therefore, this paper presents a computer-aided classification of pneumonia, coined as Ensemble Learning (EL), to simplify the diagnosis process on chest X-ray images. Our proposal is based on Convolutional Neural Network (CNN) models, which are pre-trained CNN models that have been recently employed to enhance the performance of many medical tasks instead of training CNN models from scratch. We propose to use three well-known CNN pre-trained (DenseNet169, MobileNetV2 and Vision Transformer) using the ImageNet database. Then, these models are trained on the chest X-ray data set using fine-tuning. Finally, the results are obtained by combining the extracted features from these three models during the experimental phase. The proposed EL approach outperforms other existing state-of-the-art methods, and it obtains an accuracy of 93.91% and a F1-Score of 93.88% on the testing phase.

I. INTRODUCTION

The paper motivates automated pneumonia classification on chest X-rays and proposes an ensemble learning method that combines transfer-learned CNN models and a vision transformer.

  • Pneumonia affects approximately 7.7% of the world’s population each year, making early detection important.
  • Deep learning has become widely used for medical image classification and has shown effective predictive ability on chest X-ray tasks.
  • The proposed EL method trains MobileNet, DenseNet, and Vision Transformer models using transfer learning to detect pneumonia in chest X-ray images.
  • The ensemble combines promising CNN models with a vision transformer, extending prior approaches that merged CNN models without transformers.
  • The method uses forecasts from multiple CNN models to improve classification results instead of training a CNN from scratch.
  • The proposed architecture adds batch normalization and dropout layers to the ensemble learning method.

II. RELATED WORKS

Related work shows extensive deep-learning use for pneumonia detection, while the paper distinguishes its approach by combining CNN models with a vision transformer.

  • Deep learning has been widely applied to lung-infection and disease detection from chest X-rays, including pneumonia classification and comparison with radiologists.
  • Transfer learning has become popular because it can make CNN models more efficient, reduce costs, and require fewer inputs.
  • Prior approaches used deep CNNs, transfer learning, and lung-focused regions of interest for normal-versus-pneumonia classification.
  • The proposed EL method combines three well-known CNN models, including a vision transformer, to lightly improve state-of-the-art performance with fewer layers and features.

III. METHODOLOGY

The methodology builds on deep convolutional models and ensemble techniques that extract or combine model outputs for medical-image classification.

  • Deep convolutional neural networks automatically extract features and use adjustable structures to improve machine-learning productivity and effectiveness.
  • CNN ensemble techniques either aggregate features extracted by different CNNs or merge their predicted values for classification.

1) MobileNet:

The model descriptions cover MobileNet’s efficient separable convolutions, DenseNet’s feature reuse, and Vision Transformer processing of image patches and their relationships.

  • 1) MobileNet:: MobileNet uses depthwise and pointwise separable convolutions to reduce calculation time and model size.
  • 1) MobileNet:: DenseNet links layers to improve information and gradient transfer while reducing duplicate features and parameter usage.
  • 1) MobileNet:: DenseBlocks keep feature-map dimensions consistent within a block while allowing the number of filters to vary.
  • 1) MobileNet:: Vision Transformer divides images into patches, flattens them into embeddings, and uses transformer processing for image classification.
  • 1) MobileNet:: The transformer encoder combines patch embeddings, positional embeddings, and class tokens through its attention-based architecture.
  • 1) MobileNet:: The MHSA output passes through skip connections, normalization, and an MLP block before producing the encoder-layer output.
  • 1) MobileNet:: Vision Transformer focuses on individual image patches and their relationships, unlike convolutional networks that learn features through convolutional filters.

B. Proposed EL method

The proposed ensemble method extracts medical-image representations through three pretrained deep-learning models operating in parallel, then combines their feature outputs for chest X-ray classification.

  • The ensemble method uses MobileNetV2, DenseNet169, and Vision Transformer as three parallel pretrained-model branches.Each functional layer represents one pretrained model receiving the input image.
  • The method combines feature vectors extracted by the three models to support pneumonia classification.
  • The three models are fine-tuned on chest X-ray images resized to 224 × 224 pixels after ImageNet pretraining.They generate flattened feature vectors of sizes 1280, 1664, and 768, respectively.

IV. EXPERIMENTAL STUDY

The experimental study trains and evaluates CNN-based methods and the proposed ensemble on chest X-ray classification, using regularization and a final comparison with state-of-the-art methods.

  • Nine established CNN methods and the proposed Ensemble Learning method were trained to classify pneumonia in chest X-ray images.
  • The study uses chest X-ray samples representing normal and pneumonia images for the classification task.
  • Early stopping was used because the methods began to overfit after 20 epochs.Training also used a batch size of 32, learning rate of 1e −4, Adam optimization, categorical cross-entropy loss, and softmax classification.
  • The experimental study presents the data sets and performance measures, reports experimental results, and compares the proposed method with state-of-the-art methods.

A. Data set description

The study uses a publicly available pediatric chest X-ray database containing normal and pneumonia images from patients aged one to five years.

  • The data set comes from chest X-ray scans of pediatric patients aged one to five years at Guangzhou Women and Children’s Medical Center.
  • The publicly available chest X-ray pneumonia data set contains 5,856 normal images and pneumonia images.

B. Evaluation metrics

The evaluation uses precision, recall, F1-score, and accuracy to assess classification performance, with tables reporting results for CNN models and comparisons involving the proposed ensemble.

  • The evaluation measures are precision, recall, F1-score, and accuracy.These metrics are introduced through the study’s evaluation equations for medical image classification.
  • Accuracy is defined as the system’s recognition rate and is used to evaluate baselines for each task in the two main phases.
  • Table I reports results for well-known CNN models.
  • Table II compares testing-data results for the proposed Ensemble Learning method and three well-known CNN models.

C. Results and analysis

The proposed ensemble was evaluated against MobileNetV2, DenseNet169, and Vision Transformer using training, validation, test, and confusion-matrix analyses. It achieved the strongest reported validation metrics and was described as more reliable and robust on the test set.

  • Model selection: MobileNetV2, DenseNet169, and Vision Transformer were selected because they had the best testing accuracy among the compared pre-trained models.The selected models formed the basis for the proposed ensemble learning method.
  • Model performance: MobileNetV2, DenseNet169, Vision Transformer, and Ensemble Learning were compared using training and validation losses and accuracy.The comparison is presented in Figure 3, with the ensemble showing the lowest reported validation loss and highest validation accuracy.
  • Model performance: 1.0 validation accuracy and 0.0421 validation loss were reported for the proposed ensemble method.Its training accuracy was 0.9525 and training loss was 0.1361.
  • Test-set analysis: The test-set analysis reported better accuracy and lower loss for the proposed method, describing it as more reliable and robust.The test-set accuracy comparison is referenced in Table II.
  • Classification analysis: Confusion matrices compared true and predicted labels for the four approaches in the binary chest X-ray classification task.The analysis was intended to clarify classification behavior across normal and pneumonia cases.

D. Compared Methods

The paper compares its ensemble with transfer-learning, handcrafted, GAN-based, residual, dilated-convolution, and from-scratch CNN approaches for chest X-ray pneumonia detection. The comparison emphasizes combining multiple pre-trained models while avoiding some data and training demands associated with training CNNs from scratch.

  • Comparison scope: The study compares advanced pneumonia-detection methods on the chest X-ray dataset using state-of-the-art accuracy results.Table III reports comparative test-set accuracy, with the best results labeled in bold.
  • Compared methods: Prior approaches used GAN-generated samples, transfer learning, fine-tuning, handcrafted techniques, residual connections, and dilated convolutions.The compared methods include work by Madani et al., Kermany et al., Ayan and Ünver, Stephen et al., and Liang and Zheng.
  • Comparative outcome: The proposed ensemble was reported to outperform a pre-trained CNN model in the authors’ test results.The comparison specifically discusses an automatic transfer-learning method based on DenseNet121 pre-trained concepts.
  • Training trade-offs: Training CNNs from scratch requires more data, training time, and epochs to improve generalization on input data.This is presented as a contrast with using pre-trained models and transfer learning.
  • Method limitations: The proposed method’s transfer-learning and fine-tuning hyperparameters require selection of the pre-trained model, fully connected-layer size, and number of frozen layers.The paper identifies trial-and-error or experience-based parameter selection as a drawback.

V. CONCLUSIONS

The paper concludes that its CNN ensemble identifies normal and pneumonia patients from chest X-ray images using three selected models and added classification components. It reports satisfactory classification performance and proposes a weighted ensemble as future work.

  • Conclusion: The proposed CNN Ensemble Learning method automatically identifies normal and pneumonia patients in chest X-ray images.The ensemble uses DenseNet169, MobileNetV2, and Vision Transformer selected from the trained CNN models.
  • Architecture: Global average pooling, fully connected classifier layers, and the incorporated model capabilities were reported to enhance classification performance.The global average pooling layer was merged after the convolutional layers to avoid losing spatial information.
  • Outcome: The proposed ensemble achieved satisfactory classifier performance on the chest X-ray dataset.The conclusion presents this as the overall outcome of the study.
  • Future work: Future studies will build a weighted ensemble based on the CNN models’ accuracy.This extends the current ensemble design rather than replacing the selected models.
Loading 2312.07965v1…