Multi-View Ensemble Convolutional Neural Network to Improve Classification of Pneumonia in Low Contrast Chest X-Ray Images

Abstract

Pneumonia is one of the leading causes of childhood mortality worldwide. Chest x-ray (CXR) can aid the diagnosis of pneumonia, but in the case of low contrast images, it is important to include computational tools to aid specialists. Deep learning is an alternative because it can identify patterns automatically, even in low-resolution images. We propose herein a convolutional neural network (CNN) architecture with different training strategies towards detecting pneumonia on CXRs and distinguishing its subforms of bacteria and virus. We also evaluated different image pre-processing methods to improve the classification. This study used CXRs from pediatric patients from a public pneumonia CXR dataset. The pre-processing methods evaluated were image cropping and histogram equalization. To classify the images, we adopted the VGG16 CNN and replaced its fully-connected layers with a customized multilayer perceptron. With this architecture, we proposed and evaluated four different training strategies: original CXR image (baseline), chest-cavity-cropped image (A), and histogram-equalized segmented image (B). The last strategy method (C) implemented is based on ensemble between strategies A and B. The performance was assessed by the area under the ROC curve (AUC) with 95% confidence interval (CI), accuracy, sensitivity, specificity, and F1-score. The ensemble model C yielded the highest performances: AUC of 0.97 (CI: 0.96-0.99) to classify pneumonia vs. normal, and AUC of 0.91 (CI: 0.88-0.94) to classify bacterial vs. viral cases. All models that used pre-processed images showed higher AUC than baseline, which used the original CXR image. Image cropping and histogram equalization reduced irrelevant information from the exam, enhanced contrast, and was able to identify fine CXR texture details. The proposed ensemble model increased the representation of inflammatory patterns from bacteria and viruses with few epochs to train the deep CNNs. Clinical relevance-Deep learning can identify complex radiographic patterns in low contrast images due to pneumonia and distinguish its subforms of bacteria and virus. The correlation of imaging with lab results could accelerate the adoption of complementary exams to confirm the disease's cause.

Evaluation of Scalability and Degree of Fine-Tuning of Deep Convolutional Neural Networks for COVID-19 Screening on Chest X-ray Images Using Explainable Deep-Learning Algorithm

“…Therefore, all images used in this study required contrast correction through the histogram equalization technique and resizing to a uniform size before the experiment. In this study, preprocessing was performed using the contrast limited adaptive histogram equalization (CLAHE) technique [ 25 ], which has been adopted in previous studies related to lung segmentation and pneumonia classification [ 26 , 27 , 28 ]. Figure 2 shows sample images with CXR contrast corrected using the CLAHE technique.…”

Interpretable multi-stream ensemble learning for radiographic pattern recognition

“…We herein introduced a multi-stream ensemble learning approach that could increase feature representation and identify intricate radiographic patterns, yielding higher AUCs than state-of-the-art single, limited models (p < 0.05) [2,3,5]. Moreover, the method automatically detects radiographic patterns, bringing significant benefits to the clinical routine at the beginning of care to prioritize abnormal exams for further reading from a specialist, ultimately optimizing examination time [6].…”

“…deep-learning methods were trained with single models, leading to limited prediction accuracy, even with optimum parameters [3].…”

“…Deep learning can analyze large volumes of images and automatically identify patterns accurately, including in CXR. However, most deep-learning methods were trained with single models, leading to limited prediction accuracy, even with optimum parameters [3].…”

Deep learning applied to automatic disease detection using chest X‐rays

“…Pre‐processing data also improved DNN model performance. Ferreira 40 showed that image cropping and histogram equalisation achieved AUROC 0.97 in detecting pneumonia and 0.91 in classifying the cause into bacterial or viral. Liang 41 combined residual thought (minimises overfitting/depth degradation) and dilated convolution (reduces feature space information loss) in pretrained models to detect childhood pneumonia, achieving recall rate 96.7%, and F1‐score 92.7%.…”