Research On Small Objects Detection And Segmentation In Medical Images

Computer-aided detection and segmentation of lesions has been widely used in clinical practice.In these practices,largesize objects can obtain better detection and segmentation results,but the detection and segmentation results of small objects like early tumor detection,vascular plaque segmentation,etc.are not satisfactory.The detection and segmentation of small objects in medical images has problems such as the area of small objects,few extractable features of small objects,and susceptibility to noise interference.Recently,there is little relatively research work specifically addressing these issues.Therefore,exploring how to use the current mainstream technology and making key improvements in the specific field of medical images to make it more suitable for small object detection and segmentation is an important research direction at present.There are many detection and segmentation algorithms based on deep learning,among which Mask R-CNN,as a deep learning framework that can complete detection and seg-mentation at the same time,has a wide range of applications in the field of medical images.This framework can achieve better results in detecting general large objects,but there are still problems in the detection and segmentation of small objects in medical images,such as weak feature extraction,difficulty in obtaining candidate frames,and difficulty in using small object features.This article focuses on enhancing the feature extraction capabilities of Mask R-CNN,improving the accuracy of candidate frame acquisition and improving the way of using small object features? proposes a feature extraction method based on heat maps,a candidate frame extraction scheme based on probability,and a way using edge information to segment small objects.The main innovations are as follows:1.Improve the backbone network and significantly improve feature extraction capabilities.In the detection and segmentation of small objects,stronger features the ordinary objects are needed to obtain better detection and segmentation results.In this paper,the Heat Map module is used to supervise the parameter adjustment process of the backbone network to obtain better optimization results.At the same time,the heat map generated by the module is used to amplify the signal of the feature map output by the backbone network to further enhance the extracted features,which facilitate subsequent small object detection and segmentation.Experiments show that this method can extract better features and obtain better detection and segmentation results.2.Combining medical prior knowledge,extracting object candidate frames based on probability,narrowing the target search range.Mask R-CNN complete the object detection work by uniformly presetting candidates on the image and then estimating the error between the candidate and the target.In this way,when the small object detection is performed,it is easy to cause the target fall between two adjacent candidates,making it impossible to complete the matching between the detection frame and the target frame.In this paper,a nonuniform probability based candidate frame extraction scheme is used to extract denser candidates in areas with high occurrence probability of the object to be detected,so as to improve the success rate of matching the candidate with the target.Compared with the candidate extraction scheme of Mask R-CNN,this scheme can extract candidate with smaller error with the target,and the accuracy of small object detection is also higher.3.Use boundary features to improve the accuracy of small object segmentation.When performing small object segmentation,there are fewer features available,and the Boundary information of the object will become the key feature of segmentation.This paper uses the Boundary information of the object through two modules of Boundary prediction and Boundary subdivision,thereby improving the segmentation performance of small objects.The boundary prediction module explicitly encodes the edge features of the object into the network structure,and then predicts the edge shape of the object through a special branch,which indirectly improves the segmentation result.The boundary subdivision module first generates a rough segmentation result,and then gradually subdivides the edge of the object,and finally obtains an accurate segmentation result.These two modules can be combined organically,which can significantly improve the segmentation result of small medical images.