4131

ADC and Size Dependent Segmentation Performance using Deep Learning
Chun-Jung Juan1, Yi-Jui Liu2, Shao-Chieh Lin3, and Yi-Hung Jeng4
1Department of Medical Imaging, China Medical University Hsinchu Hospital, Hsinchu, Taiwan, 2Department of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan, 3Ph.D. program in Electrical and Communication Engineering, Feng Chia University, Taichung, Taiwan, 4Department of Biomedical Engineering and Environmental Sciences, National Tsing Hua University, Hsinchu, Taiwan, Hsinchu, Taiwan

Synopsis

Accurate automatic segmentation of acute ischemic infarction on diffusion-weighted images (DWI) is clinically important. However, the accuracy of automatic segmentation of stroke lesions is affected by a lot of factors. By applying graded ADC thresholds, our study verifies the value of ADC threshold on the performance of the deep learning models in segmenting acute ischemic infarction with increasing the Dice similarity coefficient (DSC) by the lowering the ADC threshold. In addition, our study provides a new window to distinguish cytotoxic edema and vasogenic edema in acute stroke. Moreover, our results further show a size-dependent influence of DSC for stroke segmentation.

Introduction:

Although automatic segmentation has been increasingly applied to hyper-acute to acute ischemic stroke using DWI, the performance of segmentation varies widely from study to study with the Dice similarity coefficient (DSC) ranging from 0.34 [1] to 0.85 [2]. The wide variation might be attributed to factors including the different images (DWI or ADC maps) serving for ground truth segmentation [3,4], different segmentation methods (manual segmentation or thresholding segmentation) defining the ground truth, and different sizes of infarction. We hypothesized that performance of automatic segmentation is dependent on the ADC threshold defining ground truth and the size of the ischemic infarction. The aim of this study was to examine the role of ADC threshold and lesion size on segmentation performance of a deep learning method, U-net [5], for hyper-acute to acute ischemic infarction.

Materials and Methods:

This study initially enrolled 363 patients in the China Medical University Hospital (CMUH) and 120 patients in the China Medical University Hsinchu Hospital (CMUHCH) with clinical symptoms of acute ischemic infarction no more than 7 days of last known onset. MRI scans were performed using a 1.5-T scanner (Signa HDxt, GE) and a 3.0-T scanner (Signa HDx, GE). Diffusion-weighted images were acquired on axial planes with diffusion gradients (b factors) of 0 and 1000 sec/mm2 applied in each of three orthogonal directions using single shot spin-echo echoplanar acquisitions. The scan parameters included a field of view of 240 × 240 mm, a matrix of 128 × 128, a bandwidth of 651 Hz/pixel, and a section thickness of 5 mm. ADC maps were generated via pixel-by-pixel computation from b0 and b1000 images using a mono-exponential model using a formula of SIb1000=SIb0 × e-bD. Segmentation of ischemic infarction by manually contouring the hyperintense lesion on DWI was independently performed by two neuroradiologists. To evaluate the effect of ADC threshold on the performance of segmentation, an ADC threshold was added as a variable in determining the ground truth. The ADC threshold ranged from 0.6 × 10-3 mm2/sec to 1.0 × 10-3 mm2/sec at an interval of 0.1 × 10-3 mm2/sec. In the beginning, 350 original DWI images and corresponding ADC maps with stroke in CHUH were used to train artificial intelligent (AI) model for stroke lesion detection. Another 350 augmented DWI and ADC maps were generated by flipping left and right in original images. Accordingly, 700 DWI and ADC with stroke and 300 normal DWI and ADC were used for AI model training. Data were split into 80% of for training and 20% for validation. U-nets with 4 layers and 64 features were built, and 8 images as batch-size were randomly selected for each train. There were 100 interactions in each epoch, and total 150 epochs were used. Finally, a total of 6 U-net models were trained based on 6 stroke labels (original and with ADC mask 0.6 ~ 1.0 ×10-3 mm2/sec) for each observer. Fifty cases from CMUH were used for testing and 30 cases from CMUHCH were usef for externally validating the AI models for stroke lesion detection. A prediction threshold (0.8) was applied to make prediction of stroke. Dice coefficients between prediction and ground truth were calculated in 6 different types of GT to present the effect of size dependent GT by different ADC masks. ANOVA and Bouferroni post hoc analyses were applied for statistical analysis.

Result:

Figure 1 demonstrates the original b0, b1000 images and ADC map for a patient with acute ischemic stroke (a) and shows the GTs (upper column), AI predictions (middle column) and overlayed (lower column) images of the stroke lesion under different ADC thresholds ranging from 0.6 ~ 1.0 ×10-3 mm2/sec. Figure 2 demonstrates the DSCs between the GT and the prediction in different ADC thresholds and different stroke sizes. It shows a trend of higher DSC in larger stroke lesion and lower ADC threshold.

Discussion:

Our results showed significant difference across different methods in defining the ground truth and different sizes of stroke lesions. The hyperintensity on DWI might be attributed not only to the cytotoxic edema of stroke lesion [6] but also the vasogenic edema [7]. In addition, cerebral sulci included in the manually segmented GT on DWI were gradually removed while a lower ADC threshold was applied. Our results supports the hypothesis that performance of automatic segmentation is dependent on the ADC threshold defining ground truth and the size of the ischemic infarction.

Acknowledgements

No acknowledgement found.

References

  1. Perez Malla CU, Valdes Hernandez MDC, Rachmadi MF, Komura T. Evaluation of enhanced learning techniques for segmenting ischaemic stroke lesions in brain magnetic resonance perfusion images using a convolutional neural network scheme. Front Neuroinform. 2019;13:33
  2. Woo I, Lee A, Jung SC, Lee H, Kim N, Cho SJ, et al. Fully automatic segmentation of acute ischemic lesions on diffusion-weighted imaging using convolutional neural networks: Comparison with conventional algorithms. Korean J Radiol. 2019;20:1275-1284
  3. Lee H, Jung K, Kang DW, Kim N. Fully automated and real-time volumetric measurement of infarct core and penumbra in diffusion- and perfusion-weighted mri of patients with hyper-acute stroke. J Digit Imaging. 2020;33:262-272
  4. Yu Y, Xie Y, Thamm T, Gong E, Ouyang J, Huang C, et al. Use of deep learning to predict final ischemic stroke lesions from initial magnetic resonance imaging. JAMA Netw Open. 2020;3:e200772Sözen, Tümay et al. “An overview and management of osteoporosis.” European journal of rheumatology vol. 4,1 (2017): 46-56. doi:10.5152/eurjrheum.2016.048
  5. Ronneberger O FP, Brox T. . U-net: Convolutional networks for biomedical image segmentation. 18th Medical Image Computing and Computer Assisted Interventions (MICCAI). 2015
  6. Schaefer PW, Grant PE, Gonzalez RG, Diffusion-weighted MR imaging of the brain. Radiology. 2000 Nov; 217(2):331-45.
  7. Rivers CS, Wardlaw JM, Armitage PA, Bastin ME, Carpenter TK, Cvoro V, Hand PJ, Dennis MS. Persistent infarct hyperintensity on diffusion-weighted imaging late after stroke indicates heterogeneous, delayed, infarct evolution. Stroke. 2006 Jun;37(6):1418-23. doi: 10.1161/01.STR.0000221294.90068.c4. Epub 2006 Apr 27. PMID: 16645138.

Figures

(a)Stroke images and labeling region; (b) GT, Prediction and Overlap in each mask.

The DSCs between the GT and the prediction in different ADC thresholds and different stroke sizes.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
4131