Integration Of Deep Convolutional Network And Feature Pyramid Network For Multilesion Detection

Manually detecting lesions in CT images is time-consuming and laborious,and it is also a burden for timely proposing treatment.Computer-aided detection systems can assist physicians efficiently in diagnosing disease by offering lesion prediction.However,systems designed for single lesion type of medical image datasets cannot recognize other types of lesions accurately.Thus,it could not be effective to provide comprehensive detection results.Aims to the multi-lesions CT image dataset—DeepLesion,an universal lesion detection network is proposed in this paper by integrating deep convolutional neural network and feature pyramid network.The main contents are as follows:Several experiments are conducted to investigate network that performs best in the Deep Lesion dataset by combining different feature extraction networks and cascaded structures.The final adoption is a combination model of dense convolutional network（Dense Net）and bi-directional feature pyramidal network（Bi FPN）that based on Mask RCNN,and it can be divided into two parts.1、Backbone.The proposed model employs densenet-121 as backbone network and the original structure should be simplified and adjusted for necessary.To utilize 3D spatial information from CT images,network’s inputs are set as a group of CT slices with multiple adjacent slices and central slice.In the backbone network,weighted fusion will be applied to those slices by using feature fusion algorithm and squeeze and excitation module.It can enhance or weaken the influence of adjacent slices on the fused slices and improve the detection accuracy.The proposed model takes Bi FPN as a multipledimension feature extraction network to accomplish weighted features fusion and improve the performance of feature fusion.Meanwhile,Bi FPN should be adjusted to solves the incompatible dimension problem caused by the image cropping.2、Detection and segmentation branches.The proposed network adopts detection network with multi threshold and stage,enabling that samples are closer to the true labels.Meanwhile,it can resample the proposed samples to strengthen the training process.Then,samples with high quality are fed into the detection and segmentation branch for location coding and region segmentation.On the DeepLesion testing set,the proposed model has achieved up to 85.61%detection accuracy without segmentation branch,and the three-stage cascaded model has the endpoint distance error of 1.15 mm and diameter error of 1.57 mm between the segmentation prediction results and the true label.The segmentation performance is better than the previous networks’ performance.Therefore,the proposed model can meet the needs of medical personnel to detect lesion of multi-lesion CT images.