A Transfer Learning Method for Pneumonia Classification and Visualization

Abstract

Pneumonia is an infectious disease that affects the lungs and is one of the principal causes of death in children under five years old. The Chest X-ray images technique is one of the most used for diagnosing pneumonia. Several Machine Learning algorithms have been successfully used in order to provide computer-aided diagnosis by automatic classification of medical images. For its remarkable results, the Convolutional Neural Networks (models based on Deep Learning) that are widely used in Computer Vision tasks, such as classification of injuries and brain abnormalities, among others, stand out. In this paper, we present a transfer learning method that automatically classifies between 3883 chest X-ray images characterized as depicting pneumonia and 1349 labeled as normal. The proposed method uses the Xception Network pre-trained weights on ImageNet as an initialization. Our model is competitive with respect to state-of-the-art proposals. To make comparisons with other models, we have used four well-known performance measures, obtaining the following results: precision (0.84), recall (0.99), F1-score (0.91) and area under the ROC curve (0.97). These positive results allow us to consider our proposal as an alternative that can be useful in countries with a lack of equipment and specialized radiologists.

Deep-Pneumonia Framework Using Deep Learning Models Based on Chest X-Ray Images

“…Several methods have been introduced in the literature to help in detecting pneumonia using chest X-ray images. Some of these methods use handcrafted feature extraction techniques along with a machine learning algorithm as a classification technique, whereas others use deep learning techniques for feature extraction and classification [11]. These methods have changed the parameters of deep layered CNNs for pneumonia detection that can be used to obtain high accuracy in disease detection.…”

Fast COVID-19 and Pneumonia Classification Using Chest X-ray Images

“…Kermany et al [29], presented a pneumonia dataset with only 5232 images and established a baseline result for it. Additionally, on Kermany's dataset, multiple works had been presented; the three latest top results are by Lian and Zheng [28], Luján et al [30], and Chouhan et al [31]. On the RSNA Challenge [32], the top result was presented by Sirazitdinov et al [33].…”

Efficient Pediatric Pneumonia Diagnosis Using Depthwise Separable Convolutions

“…Ayan and Ünver [22] have exploited VGG16 [10] and Xception [9] for pneumonia detection in CXRs, however, their highest accuracy was only 0.87 (with low recall of 0.82). More recently, Thakur et al [38] and Luján-García et al [40] have also repurposed pre-trained models like VGG16 [10] and Xception [9] for pneumonia detection in images. The results reported in these studies [38,40] are better with much higher recall rates of 0.987 in [38] and 0.992 in [40].…”

“…More recently, Thakur et al [38] and Luján-García et al [40] have also repurposed pre-trained models like VGG16 [10] and Xception [9] for pneumonia detection in images. The results reported in these studies [38,40] are better with much higher recall rates of 0.987 in [38] and 0.992 in [40]. Recall is often considered a highly critical performance measure for a model as it gives a measure of false negatives.…”

“…All the past models [7,[20][21][22][23][24][38][39][40][41], except [43], have been evaluated in terms of only performance and not in terms of efficiency. Unlike other related work, Hu et al [43] have worked on model efficiency rather than just accuracy.…”