Semi-Supervised Learning in Medical Image Segmentation: Brain Tissue with White Matter Hyperintensity Segmentation using FLAIR image

White matter hyperintensity (WMH) has been considered the primary biomarker from 1 small-vessel cerebrovascular disease to Alzheimer’s disease (AD) and has been reported for its 2 correlation of brain structural changes. To perform WMH related analysis with brain structure, both 3 T1-weighted (T1w) and (Fluid Attenuated Inversion Recovery(FLAIR) are required. However, in a 4 clinical situation, it is limited to obtain 3D T1w and FLAIR images simultaneously. Also, the most of 5 brain segmentation technique supports 3D T1w only. Therefore, we introduced the semi-supervised 6 learning method that can perform brain segmentation using FLAIR image only. Our method achieved 7 a dice overlap score of 0.86 for brain tissue segmentation on FLAIR, with the relative volume difference 8 between T1w and FLAIR segmentation under 4.8%, which is just as reliable as the segmentation 9 done by its paired T1w image. We believe our semi-supervised learning method has a great potential 10 to be used to other MRI sequences and provide encouragement to people who seek brain tissue 11 segmentation from a non-T1w image. 12

Larger WMH are associated with an accelerated cognitive decline and increased risk for AD [4]. 23 Recent studies suggest that WMH may play a role in AD's clinical symptoms and a synergistic 24 contribution of both medial temporal lobe atrophy (MTA) and WMH on cognitive impairment 25 and dementia severity in AD [5]. Patients with mild cognitive impairment (MCI) or early AD had 26 concurrent WMH, which shows more significant cognitive dysfunction than those with a low WMH 27 burden [5]. WMH may predict conversion from MCI to AD [6]. Besides cognitive impairment, WMH 28 has reported a relationship with structural changes and cognitive performance, specially in processing 29 speed, even in cognitively unimpaired participants [7]. 30 In clinical practices, WMH burden is usually estimated by visual scale such as Fazekas' rating 31 scale, but it is difficult to use as an objective indicator. Quantification of WMH is essential to evaluate 32 the association of WMH burden with cognitive dysfunction and longitudinal change of WMH volume. 33 So, a reliable automated method for measuring WMH and cortical volume is helpful in clinical 34 practices. Recently, it is reported that WMH progression is associated with more rapid cortical thinning. 35 Therefore, the automated WMH burden and cortical volume measurement method on FLAIR is 36 clinically valuable for trace the longitudinal change in patients with cognitive impairment [8]. 37 Thus, brain structural analysis, specially volumetric analysis, combined WMH could provide more 38 descriptive information to reveal the the relationship between cognitive performance and MRI-based 39 biomarkers. 40 There are various brain tissue segmentation tools in 3D T1-weighted (T1w) images, such as 41 FreeSurfer [9], SPM [10], and FSL [11]. However, the other MR sequence (Fluid Attenuated Inversion 42 Recovery(FLAIR), Susceptibility weighted imaging(SWI), Gradient echo sequence(GRE), etc.) rarely 43 developed brain tissue segmentation because it does not intend to measure brain volume or analyze 44 brain morphology precisely. In reality, it is rare for clinicians to obtain both 3D T1w and FLAIR images 45 due to the Burden of scanning time. Ultimately, this fact hinders the approach of performing brain 46 tissue segmentation on non-T1w sequences.
In this study, we propose a brain tissue segmentation method to obtain a trainable brain label 48 on FLAIR. With given T1w and FLAIR paired datasets, we initially generate brain labels on FLAIR 49 images. Then, improved the label quality using the semi-supervised learning method. Finally, we train 50 a deep neural network-based brain segmentation model for FLAIR MRI data. The proposed method 51 could be applied in SWI, GRE, and T2, which is hard to obtain brain segmentation labels by itself, to 52 generate a trainable dataset. The following study was approved by Institutional Review Board(IRB). As shown in Table 1  Our goal is to produce a brain tissue and WMH segmentation on the FLAIR image exclusively.

71
Since the FLAIR image itself lacks the structural information and give the hardship to create the ground-truth label, we will proceed with the following process shown in Figure 1 Reference segmentation of WMH were performed by manual outlining on the FLAIR images.

83
A total of 308 datasets were manually segmented on FLAIR images, producing binary masks with

Co-registration 89
Co-registration is a method to align two individual MRIs obtained from the same subject. This 90 process is required when especially alignment needs to be made with MRIs with a different modality 91 from the same subject. In our case, this is used to align the T1w image to the FLAIR image. Since 92 the primary purpose of the Brain tissue segmentation from FLAIR is to generate initial brain tissue 93 labels on FLAIR images, we calculated the transform matrix from T1-weighted (T1w) MRI to FLAIR 94 MRI using the spatial co-registration method with SimpleITK [12]. Then we segmented T1w brain 95 tissue using T1w brain segmentation tools. We performed the brain tissue label registration on the 96 T1w image and the transform matrix based on FLAIR images' brain tissue label. The co-registrated 97 brain tissue label is on FLAIR images. However, due to differences in image spacing and dimension, it 98 did not delineate brain tissue structure accurately. Therefore, we iteratively enhanced the brain tissue 99 segmentation labels of FLAIR images.
with EvoNorm layer [14]. Histogram Equalization, Rescale-Intensity, and Z Normalization were used 104 for the pre-processing, and we set the input shape as 196 x 196.

105
With the following process, the CNN was able to learn the geometrical representation of brain 106 morphology. Thus, the delineation was more precise afterward than the brain tissue label obtained 107 T1w brain segmentation label. However, a small region distant from the brain could miss-labeled, 108 or the peri-ventricular area could be assigned as none tissue due to the initial brain tissue label's 109 impreciseness. Therefore, we performed a morphological correction (removing isolated label) to 110 generate the final brain tissue label. After training brain tissue segmentation in FLAIR image, segmentation rather delineates brain 113 tissue clearly than the brain tissue label obtained from T1w images. However, there some noise 114 that incomplete training label data could influence. Thus, we perform a simple morphological 115 correction method based on brain structure characteristics to enhance brain tissue label. All brain tissue with the T1w image and WMH. Therefore, our primary evaluation method is the dice overlap score

182
[24], which measures the similarity between the ground-truth and the prediction. In our case, the Since the dice overlap is the comparison between the reference FLAIR (ground-truth) and the 188 predicted segmentation, we will need to measure the difference between the reference T1w and the 189 predicted segmentation to see if the prediction would match the most fundamental ground-truth.

190
• Relative Difference (X, X re f erence ) = X − X re f erence X re f erence * 100 191

192
In this section, we will compare each model's detailed segmentation performance by comparing   To evaluate the average dice overlap score, we demonstrate the following Table 2, which has 199 the average dice overlap score of each model on each brain tissue. All the models exceeded the 200 average dice overlap score of 0.80 as shown in Table 2. Also, HighRes3DNet, which had an entire 3D    Table 3. The volume of individual brain tissue of the reference T1, reference FLAIR, U-Net++, U-Net, and HighRes3DNet with the relative difference between the reference T1 and the designated column. GM, gray matter; SD, standard deviation; WM, white matter. Table 3 shows the volume of segmentation (ml) and the relative difference between the segmented 212 volume and the reference T1 (%). The volume of individual brain tissue shared the similarity in   value of precision and recall were higher than the dice overlap score itself as shown in Table 4.

221
In this study, we developed reliable automated segmentation method using FLAIR for WMH and 222 cortical volume without 3D T1-weighted volume images. In clinical practices, it is not easy to get 3D 223 T1-weighted volume images due to long scan time, MR machine performance, and patient's condition.

224
On the other hand, FLAIR images is more common and essential sequence to evaluate the gaining 225 brain and easy to get in routine practices, so this method is applicable to more patients.

226
The result of Brain Tissue Segmentation Enhancement suggests that semi-supervised learning 227 enabled the direct training of brain tissue segmentation on FLAIR image only. By following the 228 Preprints (www.preprints.org) | NOT PEER-REVIEWED