Mujahid Muhammad, Rustam Furqan, Álvarez Roberto, Luis Vidal Mazón Juan, Díez Isabel de la Torre, Ashraf Imran
Department of Computer Science, Khwaja Fareed University of Engineering and Information Technology, Rahim Yar Khan 64200, Pakistan.
Department of Software Engineering, University of Management and Technology, Lahore 54770, Pakistan.
Diagnostics (Basel). 2022 May 21;12(5):1280. doi: 10.3390/diagnostics12051280.
Pneumonia is one of the leading causes of death in both infants and elderly people, with approximately 4 million deaths each year. It may be a virus, bacterial, or fungal, depending on the contagious pathogen that damages the lung's tiny air sacs (alveoli). Patients with underlying disorders such as asthma, a weakened immune system, hospitalized babies, and older persons on ventilators are all at risk, particularly if pneumonia is not detected early. Despite the existing approaches for its diagnosis, low accuracy and efficiency require further research for more accurate systems. This study is a similar endeavor for the detection of pneumonia by the use of X-ray images. The dataset is preprocessed to make it suitable for transfer learning tasks. Different pre-trained convolutional neural network (CNN) variants are utilized, including VGG16, Inception-v3, and ResNet50. Ensembles are made by incorporating CNN with Inception-V3, VGG-16, and ResNet50. Besides the common evaluation metrics, the performance of the pre-trained and ensemble deep learning models is measured with Cohen's kappa as well as the area under the curve (AUC). Experimental results show that Inception-V3 with CNN attained the highest accuracy and recall score of 99.29% and 99.73%, respectively.
肺炎是婴儿和老年人的主要死因之一，每年约有400万人死亡。根据损害肺部微小气囊（肺泡）的传染性病原体不同，肺炎可能由病毒、细菌或真菌引起。患有潜在疾病的患者，如哮喘患者、免疫系统较弱者、住院婴儿以及使用呼吸机的老年人，都有感染风险，尤其是在肺炎未被早期发现的情况下。尽管现有肺炎诊断方法存在，但低准确性和效率仍需要进一步研究以开发更精确的系统。本研究是一项利用X射线图像检测肺炎的类似尝试。对数据集进行预处理，使其适合迁移学习任务。使用了不同的预训练卷积神经网络（CNN）变体，包括VGG16、Inception-v3和ResNet50。通过将CNN与Inception-V3、VGG-16和ResNet50相结合构建集成模型。除了常见的评估指标外，还使用科恩kappa系数以及曲线下面积（AUC）来衡量预训练和集成深度学习模型的性能。实验结果表明，CNN与Inception-V3结合的模型分别达到了最高准确率99.29%和召回率99.73%。
