College of Engineering, Huaqiao University, Quanzhou 362021, China.
The 910th Hospital of the Joint Support Force of the Chinese People's Liberation Army, Quanzhou 362008, China.
Comput Math Methods Med. 2022 Jul 18;2022:8415187. doi: 10.1155/2022/8415187. eCollection 2022.
Pneumonia infection is the leading cause of death in young children. The commonly used pneumonia detection method is that doctors diagnose through chest X-ray, and external factors easily interfere with the results. Assisting doctors in diagnosing pneumonia in patients based on deep learning methods can effectively eliminate similar problems. However, the complex network structure and redundant parameters of deep neural networks and the limited storage and computing resources of clinical medical hardware devices make it difficult for this method to use widely in clinical practice. Therefore, this paper studies a lightweight pneumonia classification network, CPGResNet50 (ResNet50 with custom channel pruning and ghost methods), based on ResNet50 pruning and compression to better meet the application requirements of clinical pneumonia auxiliary diagnosis with high precision and low memory. First, based on the hierarchical channel pruning method, the channel after the convolutional layer in the bottleneck part of the backbone network layer is used as the pruning object, and the pruning operation is performed after its normalization to obtain a network model with a high compression ratio. Second, the pruned convolutional layers are decomposed into original convolutions and cheap convolutions using the optimized convolution method. The feature maps generated by the two convolution parts are combined as the input to the next convolutional layer. Further, we conducted many experiments using pneumonia X-ray medical image data. The results show that the proposed method reduces the number of parameters of the ResNet50 network model from 23.7 M to 3.455 M when the pruning rate is 90%, a reduction is more than 85%, FIOPs dropped from 4.12G to 523.09 M, and the speed increased by more than 85%. The model training accuracy error remained within 1%. Therefore, the proposed method has a good performance in the auxiliary diagnosis of pneumonia and obtained good experimental results.
肺炎感染是导致儿童死亡的主要原因。常用的肺炎检测方法是医生通过胸部 X 光进行诊断，并且外部因素容易干扰结果。基于深度学习方法辅助医生对患者进行肺炎诊断可以有效消除类似问题。然而，深度神经网络的复杂网络结构和冗余参数以及临床医疗硬件设备的有限存储和计算资源使得这种方法难以广泛应用于临床实践。因此，本文研究了一种基于 ResNet50 剪枝和压缩的轻量级肺炎分类网络 CPGResNet50（具有自定义通道剪枝和幽灵方法的 ResNet50），以更好地满足临床肺炎辅助诊断的高精度和低内存应用要求。首先，基于分层通道剪枝方法，以骨干网络层的瓶颈部分中的卷积层之后的通道作为剪枝对象，并在对其进行归一化后执行剪枝操作，从而获得具有高压缩比的网络模型。其次，使用优化的卷积方法将剪枝的卷积层分解为原始卷积和廉价卷积。将这两部分卷积生成的特征图组合作为下一个卷积层的输入。此外，我们使用肺炎 X 射线医学图像数据进行了许多实验。结果表明，所提出的方法在剪枝率为 90%时将 ResNet50 网络模型的参数数量从 23.7M 减少到 3.455M，减少了 85%以上，FIOPs 从 4.12G 降低到 523.09M，速度提高了 85%以上。模型训练的准确率误差保持在 1%以内。因此，所提出的方法在肺炎辅助诊断中具有良好的性能，并获得了良好的实验结果。
