Research On Universal Lesion Detection Of CT Images Based On Deep Learning

With the rapid development of medical imaging technology and computer vision,computer-aided detection and diagnosis systems(CADs)have become an important tool for doctors in clinical diagnosis,which are widely used in the field of medical imaging represented by Computed Tomography(CT),Magnetic Resonance Imaging(MRI),and Ultrasound Imaging(UI),etc.Among them,CT images are widely used in clinical practice,and are suitable for clinical diagnosis and condition assessment of various diseases.Lesion detection of CT images has become a hot research topic in the field of computer vision.At present,most of the works focus on the detection of a single type of lesion,which is not applicable for detecting other type of lesions.With the ability of detecting different type of lesions in CT images from various parts of the patient’s body,Universal Lesion Detection(ULD)of CT images is drawing more attention.This article mainly focuses on universal lesion detection of CT images.The work flow and research results are as follows:Based on the Mask R-CNN,an end-to-end universal lesion detection network is designed.Combining the domain knowledge of medical CT images and the practical experience of doctors in diagnosis,a Local Axial Scale-attention(LASA)module is designed to enhance the feature differences between lesion regions and non-lesion regions,which can solve the problem of high similarity between lesion regions and non-lesion regions in the CT image.This module provides multiple local square areas with different scales for each pixel in the input feature map,and merges features from the square areas to obtain weight value of each pixel.In order to reduce the computational complexity in the LASA module,the height-and width-axis passing through the center of the local square are used instead of the local square to provide local information.The proposed universal lesion detection framework can detect lesions effectively on the Deep Lesion dataset.The sensitivity is 92.75% when FPs is4,and the average sensitivity is 85.97% when FPs are 0.5,1,2 and 4,which outperforms many advanced universal lesion detection algorithms.In order to further improve the accuracy of the weight calculation and reduce the cumulative error in the LASA module,this thesis embeds position information to each pixel participating in the weight calculation to enrich the spatial feature.Specifically,the initial encodings of the position information are only related to relative position,which results in shared position information,so that few model parameters are introduced.Experimental results have shown that embedding position information into model can significantly improve the accuracy.Compared with the previous advanced algorithm MULAN,the sensitivity is 2.18% and 1.27% higher when FPs is0.5 and 1 respectively,and the average sensitivity is 1.34% higher when FPs are 0.5,1,2 and 4.The experimental results show that the proposed algorithm has significant benefits in terms of generalization and efficacy in universal lesion detection tasks.