基于MobileNetV2的小肠溃疡病灶图像识别

Image Recognition of Small Intestinal Ulcer Based on MobileNetV2

小肠镜下的溃疡病变形态复杂,鉴别诊断困难。为实现小肠溃疡病变的人工智能辅助识别,提高诊断效率和准确度,构建了一种基于MobileNetV2网络的小肠溃疡性病灶识别算法。以MobileNetV2为主干特征提取网络,将输出特征图进行空间上的多尺度提取后输入至通道注意力模块中进行特征重标定,并将多个尺度上的特征进行融合后输出分类。为了缓解数据集不均衡所带来的影响,提出了一种改进的损失函数。所用数据集来自上海长海医院282位患者的共2124张小肠镜临床图像。采用所提方法对该数据集测试的识别准确率为87.86%,5折交叉验证平均准确率为87.27%。使用梯度加权类激活图进行了可视化验证,同时将所提模块应用在不同主干网络架构上,均具有良好的通用性。研究表明,该网络模型能够更加注重病灶信息,加强病灶特征判别指向,对于小肠溃疡图像具有较高的识别准确率,可初步实现小肠溃疡病灶的自动识别。

Ulcer lesions under enteroscopy are complex in shape and difficult to differentiate and diagnose.To realize the artificial intelligence-assisted recognition of small intestinal ulcer lesions and improve the diagnosis efficiency and accuracy,a small intestinal ulcer lesion recognition algorithm based on the MobileNetV2 network was constructed.The MobileNetV2 was used as the backbone feature extraction network,and the output feature map was extracted in space at multiple scales and then input to the channel attention module for feature recalibration,and the features on multiple scales were fused and output classification,in order to alleviate the impact of data imbalance,an improved loss function was proposed.The data set used were collected from a total of 2124 enteroscopy clinical images of 282 patients in Shanghai Changhai Hospital.The proposed method was used to test the data set,and the recognition accuracy was 87.86%,the average accuracy of 5-fold cross-validation was 87.24%.The gradient weighted class activation map was used for visual verification.At the same time,the proposed modules were applied to different backbone network architectures,which reached an improvement to a certain extent,and displayed good versatility.Experimental results showed that the network model extracted more information of lesion,strengthened the identification of lesion characteristics,and had a higher recognition accuracy for small intestinal ulcer images,and was able to initially realize the automatic identification of small intestinal ulcer types.