Source-linked AI summary
Automated Methods for Detection and Classification Pneumonia based on X-Ray Images Using Deep Learning
Khalid El Asnaoui, Youness Chawki, Ali Idri
TL;DR
Pneumonia imaging diagnosis demands timely analysis, motivating evaluation of fine-tuned deep convolutional architectures. The paper compares these models for binary classification on chest X-ray and CT images, with several architectures achieving accuracy above 96%.
Problem
Increasing patient numbers make timely radiologist diagnosis difficult, motivating improved medical-image analysis for pneumonia.
Method
The paper fine-tunes eight deep-learning architectures for binary classification using X-ray inputs, regularization, and benchmark evaluation metrics.
Results
96.61% accuracy was achieved by Resnet50, while MobileNet_V2 and Inception_ResNet_V2 achieved 96.27% and 96.09%, respectively, outperforming lower-scoring alternatives.
Takeaways & Limitations
Fine-tuned Resnet50, MobileNet_V2, and Inception_ResNet_V2 showed highly satisfactory performance for the reported pneumonia-classification experiments.
Abstract
from arXiv · showhide
Recently, researchers, specialists, and companies around the world are rolling out deep learning and image processing-based systems that can fastly process hundreds of X-Ray and computed tomography (CT) images to accelerate the diagnosis of pneumonia such as SARS, COVID-19, and aid in its containment. Medical images analysis is one of the most promising research areas, it provides facilities for diagnosis and making decisions of a number of diseases such as MERS, COVID-19. In this paper, we present a comparison of recent Deep Convolutional Neural Network (DCNN) architectures for automatic binary classification of pneumonia images based fined tuned versions of (VGG16, VGG19, DenseNet201, Inception_ResNet_V2, Inception_V3, Resnet50, MobileNet_V2 and Xception). The proposed work has been tested using chest X-Ray & CT dataset which contains 5856 images (4273 pneumonia and 1583 normal). As result we can conclude that fine-tuned version of Resnet50, MobileNet_V2 and Inception_Resnet_V2 show highly satisfactory performance with rate of increase in training and validation accuracy (more than 96% of accuracy). Unlike CNN, Xception, VGG16, VGG19, Inception_V3 and DenseNet201 display low performance (more than 84% accuracy).
II. Related works
Related work applies deep learning and image processing to pneumonia detection, localization, and prediction, while motivating comparison of DCNN architectures for X-ray classification.
- Prior studies used convolutional and machine-learning methods to classify pneumonia from lung X-ray images.
- Residual structures and dilated convolutions were combined to address overfitting, degradation, and feature-space information loss in child-pneumonia classification.
- Mask-RCNN, RetinaNet, and Mask R-CNN were applied to identify or localize pneumonia in chest radiographs.
- Deep learning models were also evaluated for post-stroke pneumonia prediction using electronic health-record temporal information.
- The paper frames a central question: which deep-learning architecture is most suitable for automatic binary classification of X-ray images.
III. Proposed contributions
The proposed contribution fine-tunes several CNN architectures for binary classification of chest X-ray and CT images, using regularization to reduce overfitting and a defined baseline CNN.
- CNNs are trained end-to-end to extract and select image features before classification or prediction.
- III.1. Proposed baseline CNN architecture: The baseline CNN processes 244x244 X-ray inputs through convolutional, max-pooling, ReLU, fully connected, and output layers.
III.2. Deep Learning architectures
The compared architectures differ in depth, connectivity, filter structure, and convolutional operations, providing varied designs for pneumonia image classification.
- VGG architectures: VGG16 and VGG19 use simple 3×3 convolution kernels and 2×2 max-pooling layers, with VGG19 adding one layer to each convolutional block.
- Inception architectures: Inception_V3 concatenates parallel 1x1, 3x3, and 5x5 convolution outputs with 3×3 max-pooling outputs.
- Residual architectures: Resnet50 uses residual layers organized into convolution and identity blocks for image classification.
- Residual architectures: Inception-ResNet-v2 combines inception modules with residual connections and produces estimated class probabilities.
- Dense connectivity: DenseNet201 connects each layer to every other layer in a feed-forward fashion, promoting feature reuse and propagation.
- Efficient architectures: MobileNet_V2 uses depthwise separable convolutions to reduce memory and parameter requirements, forming a small and efficient model.
IV. Experimental results and analysis
The experiments use a three-step pipeline—data acquisition, preprocessing, and classification—on a publicly available chest X-ray and CT dataset illustrated with pneumonia examples.
- The proposed workflow is divided into data acquisition, data pre-processing, and classification.
- IV.1. Dataset: The dataset contains 5856 JPEG images divided into 4273 pneumonia and 1583 normal cases.
- IV.1. Dataset: Figure 4 includes normal lungs without abnormal opacification and pneumonia examples showing focal lobar consolidation.
IV.2. Data pre-processing and data splitting
Input images are pre-processed with intensity normalization and CLAHE, then split into separate training and validation sets while addressing class imbalance through resampling.
- Data pre-processing: Intensity normalization and CLAHE are applied to improve input-image quality before model processing.The stated goals include reducing noise, increasing contrast, and removing low or high frequencies.
- Data pre-processing: Min-max normalization maps input images toward a standard normal distribution before model ingestion.The paper associates this operation with Equation 1 and Figure 5.
- Data pre-processing: Figure 5 contrasts the original image, normalized image, and CLAHE-processed image.
- Data splitting: 60% of images are assigned to training and 40% to validation, with validation images excluded from training.
- Data splitting: Because 75% of the dataset belongs to the pneumonia class, resampling through data augmentation is used to address imbalance.
IV.3. Data augmentation
Data augmentation is applied after preprocessing and splitting to enlarge the training set and reduce overfitting risk using geometric image transformations.
- IV.3. Data augmentation: Data augmentation is used during training to increase dataset size and reduce the risk of overfitting.
- IV.3. Data augmentation: The augmentation strategies include rescaling, rotations, shifts, shears, zooms, and flips.These strategies are listed as the augmentation parameters in Table 2.
IV.4. Training and classification dataset
After preprocessing, splitting, and augmentation, the expanded training data is passed through proposed models for feature extraction and multilayer-perceptron classification.
- IV.4. Training and classification dataset: Proposed models extract pertinent features from the expanded training dataset after preprocessing, splitting, and augmentation.
- IV.4. Training and classification dataset: Extracted features are flattened together into vectorized feature maps before classification.
- IV.4. Training and classification dataset: A multilayer perceptron receives each feature vector and assigns the image to its corresponding class.
IV.5. Experimental setup
The experimental setup resizes images by model requirements and trains the models with fixed batch-size, epoch, optimizer, and dataset-splitting parameters.
- IV.5. Experimental setup: All images are resized to 224x224 pixels, except Inception_V3 inputs, which use 299x299 pixels.
- IV.5. Experimental setup: Training uses a batch size of 32 and 300 epochs.
- IV.5. Experimental setup: The training and validation samples are respectively set to 159 and 109.
IV.6. Evaluation criteria
The study evaluates classification performance using accuracy, sensitivity, specificity, precision, and F1 score, with confusion-matrix terms defining the underlying outcomes.
- The evaluation uses accuracy, sensitivity, specificity, precision, and F1 score as benchmark metrics for classification performance.
- TP denotes true positives, FP false positives, TN true negatives, and FN false negatives.
IV.7. Results and discussion
The results section evaluates baseline and fine-tuned deep-learning architectures for binary classification of chest X-Ray and CT images using training and validation curves.
- The study compares baseline CNN with fine-tuned VGG16, VGG19, Inception_V3, Xception, Resnet50, Inception_Resnet_V2, DenseNet201, and MobileNet_V2 architectures.
- Several experiments on chest X-Ray data assess the performance and robustness of each proposed model.
- Training and validation curves present the results separately across Figures 6–15.
IV.7.1. Classification results of the proposed architectures
The fine-tuned architectures generally reached stable training and validation behavior, but their classification performance varied substantially. ResNet50, Inception_ResNet_V2, and MobileNet_V2 were among the strongest models, while baseline CNN, Xception, VGG16, and VGG19 performed less well.
- Training and validation accuracy curves increased toward stable values across the evaluated architectures, while corresponding loss curves generally decreased.The paper uses these curves to assess learning and generalization on training and hold-out validation data.
- 96.23% test accuracy was reached by Resnet50, which correctly classified 1703 Normal and 1638 Pneumonia images.
- 96.41% test accuracy was reached by Inception_ResNet_V2, with 1618 Pneumonia images and 1705 Normal images correctly classified.
- 69.03% test accuracy at epoch 300 was reported for one evaluated architecture, with 510 Pneumonia images classified as Normal.
- 84.18% accuracy was obtained with baseline CNN, compared with 83.14% for Xception, 86.26% for VGG16, and 85.94% for VGG19.
- 96.61% accuracy was reported for Resnet50, alongside 96.27% for MobileNet_V2 and 96.09% for Inception_Resnet_V2.
V. Conclusions & Future works
The paper compares nine deep-learning architectures for binary chest X-ray and CT classification, finding the strongest performance from Resnet50, MobileNet_V2, and Inception_ResNet_V2. Future work proposes expanding the system and improving performance with additional data and advanced detection and segmentation methods.
- Future work aims to develop a complete pneumonia system covering detection, segmentation, and classification.
- Additional datasets and advanced deep-learning feature-extraction methods such as YOLO and U-Net are proposed to improve performance.