多感受野的轻量化YOLOv4用于检测肺结核

Lightweight YOLOv4 with multi-receptive fields for detection of pulmonary tuberculosis

肺结核疾病特征错综复杂,人工筛查成本较高,缺少规范的数据集。当前基于卷积神经网络的检测模型结构复杂、参数量大且检测精度有待进一步提高,为此提出一种改进的轻量化YOLOv4的肺结核检测模型。首先选取300例实际病例,制作一套规范的数据集,用于评估模型的性能;随后通过残差通道注意力模块改进MobileNetv3的结构,并作为YOLOv4的主干提取器,进一步减少参数量并融合上下文信息;然后在主干提取器的3个有效特征层后加入多感受野模块,有效增强低特征层的信息提取能力并降低对小型肺结核病灶的漏检率;最后,将以上改进的模块与YOLOv4的多尺度结构相结合,构建一种多感受野的轻量化YOLOv4的肺结核检测模型。与原始YOLOv4相比,该模型的参数量减少了约47%,平均精准度(mAP)值提升至96.60%,漏检率降低至6%,验证该模型能有效辅助影像科医师诊断肺结核。

The characteristics of pulmonary tuberculosis are complex,with the high cost of manual screening,lack of standardized data sets.The current detection model based on convolution neural network has intricate structure,large number of parameters and detection accuracy needs to be further ameliorated.Therefore,an improved lightweight YOLOv4 model is proposed for pulmonary tuberculosis detection.A standardized dataset is constructed using 300 actual cases for evaluating the performance of the model.MobileNetv3 improved with residual channel attention module is used as the backbone extractor of YOLOv4 for further decreasing the number of parameters and fusing context information.Then the multi-receptive field module is added after the 3 effective feature layers of the backbone extractor,which effectively enhances the information extraction ability of the low feature layer and reduces the missed etection rate of small pulmonary tuberculosis lesions.The above improved modules were combined with the multi-scale structure of YOLOv4 to construct a lightweight YOLOv4model with multi-receptive field for pulmonary tuberculosis detection.Compared with the original YOLOv4,the proposed model reduces the number of parameters of the model by about 47%,elevates the mAP value to 96.60%,and decreases the missed detection rate to 6%.It is verified that lightweight YOLOv4 with multi-receptive fields can effectively assist radiologists in the diagnosis of pulmonary tuberculosis.