An In-depth Research On The Segmentation Of Small Diffuse Lesions In Medical Brain Images

In recent years deep learning algorithms have been widely used in brain lesion segmentation on medical images,but unsolved problems exist at present: Deep learning has achieved 85% or even more than 90% of the lesions segmentation accuracy when segmenting focal and/or big lesions such as brain tumor or acute hemorrhage lesions at their middle and late stage,this accuracy approximates or even slightly exceeds the segmentation accuracy of medical experts.However,for small and diffuse lesions such as many diseases at their early stage,various studies commonly reported 30%-70% of the lesion segmentation accuracy.This gap in segmentation accuracy seriously hinders the detection of all kinds of small and diffuse brain injury using deep learning algorithms.It has become a technical difficulty of deep learning algorithm in the field of medical image.This paper aims to study this problem systematically.Identifying and segmenting any size of lesions is important,and it plays an important role in the clinical prognosis and assessment.However,up to now,the segmentation of small diffuse lesions in medical images still faces great challenges,the problems mainly include the following aspects:(1)Currently,there is no widely accepted definition of small diffuse lesions,and the classification of small diffuse lesions differs greatly in different pieces of literature;(2)The reasons behind the gap between the small diffuse lesion segmentation results and the state-of-the-art segmentation results are not clear;(3)The existing solutions to improve the segmentation accuracy of small diffuse lesions have not been summarized;(4)The generality of algorithms across lesion types has rarely been explored,and the evaluation criteria of small diffuse lesions' segmentation are not completely reasonable;(5)The future direction of small diffuse lesions segmentation is not clear yet.Focusing on the segmentation of small diffuse lesions in medical images,the main innovation of this paper is summarized as follows:(1)This paper explored the definition of small diffuse lesions in the current literature for the first time.This paper proposes that if the lesions are diffuse(multifocal)in nature and occupy <1% of the whole brain volume in at least 1/3 of the patients having this type of lesions,they are called small diffuse lesions(Chapter 1.2).(2)For the first time,the reasons behind the gap between the small diffuse lesion segmentation results and the state-of-the-art segmentation results were analyzed.We proposed three special challenges faced by the segmentation of small diffuse lesions(Chapter 3.1);For the first time,the improvement methods of the existing deep learning algorithms for the segmentation of small diffuse lesions are extracted,and based on the special challenges mentioned before,we analyzed how they are respectively solved by the improvement methods(Chapter 3.2).The three special challenges include a)more subtle lesion signals and small differences between lesions and non-lesions;b)localized small lesion regions,it is not suitable for deep neural network and Dice loss function;c)class imbalanced problem.(3)One kind of small diffuse lesion,hypoxic ischemic encephalopathy(HIE),is taken as an example,to verify the overlap between the lesions segmented by the proposed algorithm and the expert.The proposed method is changing the input of the neural network and adding quantified normal spatial anatomical location information to it.The median sensitivity of lesion segmentation is 0.21 if the deep learning method is used directly,and 0.64 if using the method proposed in this paper.If our results(N=133)were compared with those in the most advanced literature(N=20)in the case of similar lesion volume,our results(Dice = 0.77)were greatly improved compared with those in the most advanced literature(Dice = 0.52)(Chapter 4).(4)Without relying on the data labeled by experts,the universality of the algorithm to the type of lesions and to the diagnosis was verified by taking two kinds of small diffuse lesions as examples.We take HIE and SWS as examples,for the problem that the evaluation standard of segmentation result about small diffuse lesions is not completely reasonable,the experimental results verify that our algorithm based on quantifying normal spatial information can repeat doctors' diagnoses in different encephalopathy,and can also remedy doctors' misdiagnosis to a large extent(Chapter5).In the end,the prospects were put forward.First,by summarizing the existing public datasets of small diffuse lesions,the current big data project/research involving small diffuse lesions,and the public software for segmentation of small diffuse lesions,this paper put forward the prospect of exploring benchmark methods for cross-lesion type research based on big data.Then,the prospect of the deep learning segmentation algorithm for brain small diffuse lesions was presented(Chapter 6).