Multi-scale Object Equalization Learning For Intracerebral Hemorrhage Segmentation In CT Images

Intracerebral hemorrhage usually refers to the hemorrhage caused by intracranial cere-brovascular diseases.Computed tomography(CT)is the first choice for the initial diagnosis of intracerebral hemorrhage due to its economy,rapid and wider use.Therefore,accurately and promptly detection and segmentation of the intracerebral hemorrhage region from CT images can assist doctors to grasp the situation of the intracerebral hemorrhage region more objectively and formulate a better treatment plan.However,the manual segmentation of the intracerebral hemorrhage region by the doctor not only requires richer experience and is extremely time-consuming,but also can easily lead to the wrong segmentation due to long-term fatigue work.Moreover,a large number of clinical cases also increase the workload of doctors.After several years of development,computer-aided diagnosis(CAD)technology has been able to help doctors make diagnoses and treatments more quickly and accurately in some clinical diagnosis situations.However,due to the diversity of clinical factors and individual differences in clinical,the regions of hemorrhage show greater characteristic differences in clinical CT images.Among them,the characteristic manifestations of multi-scale objects not only exist in a single CT scan slice,but also widely exist in different CT scan slices.This undoubtedly brings great learning challenges to segmentation methods of CAD.Traditional segmentation methods require human involvement in extracting different features for segmentation.Such methods do not only require more human knowledge inter-vention,but also perform poorly in robustness.In recent years,deep learning methods have made a figure in many fields due to their powerful automatic feature learning and expres-sion capabilities.But their performance in the field of intracerebral hemorrhage remains to be further explored.With this in mind,we propose a multi-scale objects equalization learning convolutional neural network to alleviate the difficulty in equalization learning for characteristics due to the problem of large differences in clinical intracerebral hemorrhage.The main works are described as follows:1)We develop a novel MOEL-Net for multi-scale intracerebral hemorrhage objects in CT images,which more comprehensively considers the needs of different semantic abstrac-tion levels of segmentation object at different scales.A shallow feature extraction module(SFEM)that prefers small object learning is used to extract shallow semantic features that could appropriately express the information of small objects.At the same time,a deep fea-ture extraction module(DFEM)that prefers larger object learning is used to extract deeper semantic information.In particular,considering the features of larger objects are abundant,to better learn and express the deeper semantic features,we adopt a progressive strategy in DFEM to achieve multi-level semantic information extraction.Then,transferring the ex-tracted multi-level semantic features to the decoder part of DFEM by skip connect,which greatly improves the feature extraction and expression ability of DFEM.Finally,Through the multi-level semantic feature equalization fusion module(MSFEFM),the features of dif-ferent semantic abstraction levels extracted by the SFEM and DFEM modules are learned in an equalization way to achieve fine segmentation of objects with different scales.2)We collected CT scan images of clinical intracerebral hemorrhage from several hospitals,and constructed two data sets base on the composition characteristics of the case data,named VMICH and FRICH respectively.Among them,VMICH contains 460 CT cases,each of which contains only the first or review CT scan images.Such data compo-sition makes the data environment more complicated and also brings greater challenges to the segmentation algorithm.There are 480 cases of FRICH,each case includes the first and review CT images.3)In this paper,a large number of experiments conduct on clinical data sets and com-pare the MOEL-Net with the 11 mainstream segmentation networks to verify the perfor-mance of the proposed architecture.The final experiment results show that the performance of our method is superior to the 11 mainstream segmentation networks.Our method is more effective in solving multi-scale object equalization learning problems of the intracerebral hemorrhage region with unbalanced data situations to achieve accurate segmentation.