本研究目的為利用卷積神經網路的混合模型來探討卷積層深度與卷積核數量對肺炎診斷準確度的影響。在模型設計方面,本研究基於AlexNet模型架構,並套用VGGNet模型對深度與數量的假設,設計出4種不同權重的卷積層深度,每種下面分別含有5組由不同數量但由少到多疊成的卷積核架構,並且以2*2池化與3*3重疊池化區分,最後共有35組模型架構,並將其應用於5,863張胸部X光影像以診斷肺炎。研究結果顯示,3*3重疊池化中,卷積層深度為10層權重,卷積核數量為128、256、512的組合,普遍優於其他組,在準確率(86.07%)、精確率(83.74%)與綜合評估指標(89.66%)獲得最佳的結果,同時此組模型的表現亦皆優於VGGNet。結論顯示儘管過去研究證實VGGNet提出增加深度與數量可以提高模型表現的論點,但透過本研究的混合模型,發現能夠以較少的深度與數量,即可有效的提高模型表現。本研究將能協助研究者在有限的硬體資源下達到更佳的成果。
Pneumonia is a common disease and a leading cause of death in the world. Chest x ray is the most commonly performed radiologic procedure for pneumonia diagnosis. However, reading x-ray images can be complicated due to various medical conditions and a requirement for domain expertise and experience. Major convolutional neural network (CNN) model, such as VGGNet, had been proposed to detect x-ray images. However, CNN models often include tens to hundreds of millions of parameters, which produce heavy computation and memory loading and limit the practical usage in training and optimizing for real-world applications. To speed up diagnosis, we proposed a mixed CNN model to enhance the accuracy in detecting pneumonia from chest radiographs. This study developed a model, in which the AlexNet architecture and design concept of the convolution layers and filters in VGGNet are incorporated. We used four different weights (8-11) of convolution layers and five different combinations of the convolution filters and different max-pooling layers (2*2 and 3*3 overlapping) to form a total of thirty-five model architectures. Then, we applied those architectures to 5,863 chest X-ray images to diagnose pneumonia. The results indicated that the number of the (128、256、512) convolution filters with ten weights of convolution layer and 3*3 overlapping pooling had the greatest accuracy (86.07%) and precision (83.74%) and F1 measure (89.66%) among all the models. This model performance was also greater than the models of VGGNet. Although previous studies proved that the theory from VGGNet about deeper layers and more filters can improve the model performance. Through the proposed mixed model, we found that the model performance can be improved with fewer layers and filters. Future research may apply the proposed model to similar cases with constrained hardware resources.