Highly Accurate Brain Stroke Diagnostic System and Generative Lesion Model Junghwan Cho, Ph.D. CAIDE Systems, Inc. Deep Learning R&D Team
Established in September, 2016 at 110 Canal st. Lowell, MA 01852, USA CEO & Founder: Jacob Kyewook Lee Employees : 6 Contact: caideinfo@caidesystems.com Our Mission http://www.caidesystems.com Save human lives by developing C ognitive A rtificial I ntelligence D isease De tection Systems . Provide protection of human life and equal access to health care and treatment through artificial intelligence technology. Our Goals Eliminating human errors and reducing delayed diagnosis Developing the most reliable AI system for analyzing images (ultra sound, MRI, CT and X-ray), electronic medical records, and genome data. Available Position Looking for Talented Research Scientist or Engineer With CAIDE, Better and Healthier Life!
Outlines § C AI DE Diagnostic System for Brain Stroke § Stroke Classification/Stroke Lesion Segmentation § Stroke Lesion Generative Network § Demo- C AI DE m: Studio BSR
Brain Stroke § Brain attack/accident § Up to 2 million brain cells die every minute. § About 795,000 people suffer from stroke every year in US. § More than 137,000 people (17% of all strokes) die from the stroke, with a cost of approximately $76.3 billion. Hemorrhagic Stroke (Bleeding) Ischemic Stroke (Blood Blockage) *Source image from http://www.stroke.org/understand-stroke/what-stroke/stroke-facts With CAIDE, Better and Healthier Life!
CT Findings on Intracranial Hemorrhage Types Intraventricular (IVH) Intraparenchymal (IPH) Subarachnoid (SAH) Epidural (EDH) Subdural (SDH) IPH+IVH
CAIDE Diagnostic System CT Images DICOM Files---> Gray Scale (on Window Level/Width) Preprocessing False Negative VS Model 1 CNN1- Classifier Cascaded CNN Classifiers (Default Window) No Bleeding Hemorrhagic(bleeding) CNN2- Classifier Positive? (Bleeding) (Stroke Window) NO YES Model 2 Stroke Lesion Positive? Delineation (FCN) YES NO END
Sensitivity (Recall) vs Specificity ROC for Classification ROC for Classification 1 Sensitivity Specificity Th p >=0.2 ( 0.953 , 0983 ) True Positives Rate >> Th p >=0.5 (0.991, 0.961) 0.95 (Sensitivity) = (Specificity, Sensitivity) 0.9 Cascaded CT window for increasing sensitivity while preserving specificity 0.85 0 0.05 0.1 0.15 0.2 0.25 False Positives Rate (1- Specificity)
Default Window vs Stroke Window Setting 50/100 (WL/WW) 40/40 50/100 Default Brain Window Stroke Window Ground Truth • Narrow window width (high-contrast) • Increase detection of subtle abnormalities Turner, P. J., and G. Holdsworth. "CT stroke window settings: an unfortunate misleading misnomer?." The British journal of radiology 84, no. 1008 (2011): 1061-1066.
Training for Cascaded CNN Classifier (Bleeding or not) • Total data sets- 5,647 patients (3,000 no bleeding vs 2,647 bleeding) o 2D axial CT images with 512x512 size • 5-fold cross validation • Trained cascaded CNN model o Two different training solvers: Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (ADAM) o Scratch vs fined tuned using pre-trained model • Hardware computer: NVIDIA DGX-1 with 8 Tesla V100
Evaluation- Classification- (top 1- accuracy) – after 15 epoch % 100 ("-. ". ± &. .%) ("'. +' ± &. +%) 98 ("#. "" ± &. '%) 96 ("*. *+ ± &. ,%) 94 92 90 1 2 ADAM- 3 4 SGD- SGD- ADAM- Scratch Scratch Fine-tuned Fine-tuned
Evaluation – Cascade CT Window Increasing Sensitivity % 100 Specificity 350 # of False Negative CT Images 99 Sensitivity 300 250 98 200 150 97 100 96 50 0 50/100 1 50/100+ 2 95 50/100 50/100+ 50/100+ 50/100 1 2 3 4 (WL/WW) 40/40 (Cascaded) (WL/WW) 40/40 (Cascaded) 40/40
Outlines § C AI DE Diagnostic System for Brain Stroke § Stroke Lesion Segmentation § Stroke Lesion Generative Network § Demo- C AI DE m: Studio BSR
Encoder-Decoder Architecture - for sematic image segmentation • SegNet, U-Net, and Fully Convolutional Network (FCN) • Encoder • Decoder § Feature extraction (Convolution) § High resolution from low resolution § Dimensional reduction (Pooling) § Unpooling/up-sampling with § VGG 16 or ResNet transposed convolution (deconvolution) Source image from "Segnet”, IEEE transactions on pattern analysis and machine intelligence 39, no. 12 (2017): 2481-2495.
Fully Convolutional Network (FCN) for Stroke Lesion Segmentation VGG16 Network 2 3 1 5 4 2x upsampled SUM Pool4 prediction 2x upsampled Softmax SUM Pool3 prediction 8x upsampled FCN-8s 14 Long, Jonathan, Evan Shelhamer, and Trevor Darrell. "Fully convolutional networks for semantic segmentation." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 2015.
FCN Training • Total data sets- 2,647 patients (corresponding 33,391 well labeled images) • 5-fold cross validation • Fully convolutional network with ADAM solver using pre-trained model • NVIDA DGX-1 (8 V100 GPU) Histogram of Hemorrhagic Stroke Type 14000 12000 10000 8000 6000 4000 2000 0 IPH IVH EDH SDH SAH 1 2 3 4 5
Segmented Results by FCN8s after 50 Epoch Training IPH, SAH IVH DC IPH =.89, DC SAH =.77 DC IVH =.86
Segmented Results by FCN8s SDH: False Negative after 50 Epoch Training DC SAH =. 84 SAH, SDH SAH: False Positive DC EDH =.89, DC SDH =.54 EDH, SDH
Performance Evaluation Recall, Sensitivity Precision % % 100 100 #FP: Number of Pixels Falsely Positive Segmented DC: Dice Coefficient #FP>300 90 90 #FP>200 DC>5% #FP>100 80 80 DC>25% 70 70 DC>50% 60 60 Precision=TP/(TP+FP) Recall=TP/(TP+FN) 50 50 IPH IVH EDH SDH SAH 0 1 2 3 4 5 6 0 1 2 3 4 5 6 IPH IVH EDH SDH SAH
Outlines § C AI DE Diagnostic System for Brain Stroke § Stroke Classification/Stroke Lesion Segmentation § Stroke Lesion Generative Network § Demo- C AI DE m: Studio BSR
Generative Adversarial Networks (GANs) • Two networks competing against each other in a zero sum game The discriminator (D) : Distinguish real data from fake created by the generator ( x ) The generator (G) : Learn distribution of the data from (z) random noise, in an attempt to fool the discriminator Source Image from Source Image from https://www.slideshare.net/ckmarkohchang/generative- https://twitter.com/ch402/status/793911806494261248 adversarial-networks
Image to Image Translation - for Generating Stoke Lesion Images l Apply to map stoke lesion labels to corresponding lesion image. l Stoke lesion masks (segmented regions) - conditional input images to the Generator (G) as well as Discriminator (D) Target Lesion Generated (Stroke Lesion) Image Stoke Lesion Labels Source image from, Phillip, et al. "Image-to-image GAN for Generating Stoke Lesion Images translation with conditional adversarial networks." arXiv preprint (2017)
Training Pix2Pix-Tensorflow l Trained conditional GAN below conditions Total data set : 2,647 patients (corresponding 33,391 well labeled images) q : 80% training, 20% testing Learning parameters: q Learning rate =0.0002, L1 weight=100, and GAN weight=1.0 About 16 hour up to 200 epoch on NVIDIA Tesla V100 (1 GPU) q
Examples of Generated Fake CT Image after 200 Epoch Training Input Target Output Input Target Output SDH SAH EDH SAH SAH
Input Target Output Input Target Output IPH, SAH IVH, IPH SAH, IVH SAH SAH, IVH SAH, IVH, IPH SAH, EDH
Evaluation l In general, evaluating GANs is difficult Loss function makes it harder during training q FCN /Inception scores and Amazon Mechanical Turk (AMT by human) q l FCN scores : fake/generated images inferred by FCN l Clarity- threshold blurriness (variance of Laplacian) l Second discriminator- choose more realistic images FCN scores (DICE) vs Image Quality Dice: .981 Dice: .981 Dice: .976 Dice: .978 IPH 10 epochs 50 epochs 100 epochs 200 epochs
Evaluation – Recall, DICE l Evaluated by varying: Percentage of training data (based on patient number): 2.5, 10, 50, and 100% q Number of epochs : 10, 50, 100, and 200 epoch q DICE (FCN Scores) Recall
Evaluation- Data Augmentation Original Original+ Original+ Original+ 3xAugment (Real Data) Augment 2xAugment
CAIDE m: Studio BSR (Demo), Booth #726
Recommend
More recommend
