基于改进MobileNetV2的中耳炎影像分类诊断模型

Otitis Media Image Classification and Recognition Model Based on Improved MobileNetV2

摘要: 中耳炎诊断是有效防止耳道疾病进一步恶化的重要前提，针对现有中耳炎相关研究中存在的数据集量少、网络参数量大、模型识别精度低等问题，提出一种基于改进MobileNetV2模型的中耳炎分类方法。首先，在MobileNetV2网络的倒置残差结构中嵌入坐标注意力机制，增强网络对中耳炎影像特征细化能力；其次，设计改进注意特征融合模块替换原始特征简单相加过程，加强模型在跨通道背景下对不同尺度特征的提取能力；同时采用HardSwish激活函数替换原始ReLU6函数，提升模型鲁棒性；最后，减少模型瓶颈层通道数，简化模型结构。实验结果表明，所提出CIH-MobileNetV2模型在中耳炎数据集上的识别准确率和F1 Score达到91.05%和89.06%，相较于原始MobileNetV2模型，分别提高了2.31%和3.69%，参数量较初始模型减少了43%。与经典网络AlexNet、GoogleNet、VGG16、ResNet50、MobileNetV3、ShuffleNetV2等网络相比，有更高的识别准确率和F1值，因此，该研究所提出模型能够较好对中耳炎类型进行分类，为中耳炎诊断提供有效帮助。

Abstract: Otitis media recognition is an important prerequisite for effectively preventing the deterioration of otological diseases. Aiming at the problems of small amount of dataset, large number of network parameters, low model recognition accuracy, and excessive computational volume in the existing otitis media recognition research, we propose an otitis media recognition model based on improved MobileNetV2. First, the coordinate attention mechanism is embedded in the inverted residual structure of the MobileNetV2 network to enhance the network’s ability to refine the otitis media image features. Second, Iterative Attention Feature Fusion module is used to replace the simple summation of the original features, which strengthens the model’s capability of extracting features of different scales in the context of cross-channel. At the same time, the HardSwish activation function is used to replace the original ReLU6 function to improve the robustness of the model. Finally, the number of channels in the bottleneck layer of the model is reduced to simplify the model structure. The experimental results show that the recognition accuracy and F1 Score of the proposed CIH-MobileNetV2 model on the otitis media dataset reach 91.05% and 89.06%, which are improved by 2.31% and 3.69%, respectively. Compared with the original MobileNetV2 model, the number of parameters is reduced by 43%. Compared with the classical networks AlexNet, GoogleNet, VGG16, ResNet50, MobileNetV3, ShuffleNetV2, etc., there are higher recognition accuracy and F1 value, therefore, the proposed model of the institute is able to classify the type of otitis media better and provide an effective help for otitis media diagnosis.