Highly Accurate Brain Stroke Diagnostic System and Generative - - PowerPoint PPT Presentation

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

Save human lives by developing Cognitive Artificial Intelligence Disease Detection 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!

http://www.caidesystems.com

Outlines

§CAIDE Diagnostic System for Brain Stroke §Stroke Classification/Stroke Lesion Segmentation §Stroke Lesion Generative Network §Demo- CAIDE m: Studio BSR

With CAIDE, Better and Healthier Life!

§ 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.

*Source image from http://www.stroke.org/understand-stroke/what-stroke/stroke-facts

Brain Stroke

Ischemic Stroke (Blood Blockage) Hemorrhagic Stroke (Bleeding)

CT Findings on Intracranial Hemorrhage Types

Epidural (EDH) Subdural (SDH) Subarachnoid (SAH) Intraparenchymal (IPH) Intraventricular (IVH) IPH+IVH

CAIDE Diagnostic System

DICOM Files---> Gray Scale (on Window Level/Width) VS

Hemorrhagic(bleeding) No Bleeding

False Negative NO NO

Positive?

(Bleeding)

CT Images Preprocessing CNN1- Classifier

(Default Window)

CNN2- Classifier

(Stroke Window) Stroke Lesion Delineation (FCN) Positive?

YES YES

END

Cascaded CNN Classifiers Model 1 Model 2

Sensitivity (Recall) vs Specificity

Cascaded CT window for increasing sensitivity while preserving specificity

>>

Sensitivity Specificity

False Positives Rate (1- Specificity)

0.05 0.1 0.15 0.2 0.25 0.85 0.9 0.95 1

ROC for Classification

True Positives Rate (Sensitivity) ROC for Classification (0.991, 0.961) = (Specificity, Sensitivity) (0.953, 0983) Thp>=0.5 Thp>=0.2

Default Window vs Stroke Window Setting

50/100 (WL/WW) 40/40 Ground Truth

• Narrow window width (high-contrast)
• Increase detection of subtle abnormalities

50/100 Default Brain Window Stroke Window

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)
• 2D axial CT images with 512x512 size
• 5-fold cross validation
• Trained cascaded CNN model
• Two different training solvers: Stochastic Gradient Descent (SGD) and

Adaptive Moment Estimation (ADAM)

• 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

1 2 3 4 90 92 94 96 98 100

SGD- Scratch SGD- Fine-tuned ADAM- Scratch ADAM- Fine-tuned ("#. "" ± &. '%) ("*. *+ ± &. ,%) ("-. ". ± &. .%) ("'. +' ± &. +%)

%

Evaluation – Cascade CT Window Increasing Sensitivity

1 2 50 100 150 200 250 300 350

50/100 (WL/WW) 50/100+ 40/40 (Cascaded)

# of False Negative CT Images

%

1 2 3 4 95 96 97 98 99 100

50/100+ 40/40 50/100 50/100+ 40/40 (Cascaded)

Specificity

Sensitivity 50/100 (WL/WW)

Outlines

§CAIDE Diagnostic System for Brain Stroke §Stroke Lesion Segmentation §Stroke Lesion Generative Network §Demo- CAIDE m: Studio BSR

Encoder-Decoder Architecture

• for sematic image segmentation
• Encoder

§ Feature extraction (Convolution) § Dimensional reduction (Pooling) § VGG 16 or ResNet

• SegNet, U-Net, and Fully Convolutional Network (FCN)
• Decoder

§ High resolution from low resolution § Unpooling/up-sampling with transposed convolution (deconvolution)

Source image from "Segnet”, IEEE transactions on pattern analysis and machine intelligence 39, no. 12 (2017): 2481-2495.

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. 14

Fully Convolutional Network (FCN) for Stroke Lesion Segmentation

Pool4 prediction 2x upsampled

SUM

FCN-8s

Pool3 prediction 2x upsampled

SUM

8x upsampled Softmax 1 2 3 4 5

VGG16 Network

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)

1 2 3 4 5 2000 4000 6000 8000 10000 12000 14000

IPH IVH EDH SDH SAH Histogram of Hemorrhagic Stroke Type

IVH DCIVH=.86 IPH, SAH DCIPH=.89, DCSAH=.77

Segmented Results by FCN8s after 50 Epoch Training

SAH, SDH DCSAH=. 84 EDH, SDH DCEDH=.89, DCSDH=.54 SDH: False Negative SAH: False Positive

Segmented Results by FCN8s after 50 Epoch Training

1 2 3 4 5 6 50 60 70 80 90 100 1 2 3 4 5 6 50 60 70 80 90 100

Performance Evaluation

IPH IVH EDH SDH SAH

% %

IPH IVH EDH SDH SAH

#FP: Number of Pixels Falsely Positive Segmented #FP>300 #FP>200 #FP>100

Precision Recall, Sensitivity

DC>5% DC>25% DC>50%

Precision=TP/(TP+FP) Recall=TP/(TP+FN)

DC: Dice Coefficient

Outlines

§CAIDE Diagnostic System for Brain Stroke §Stroke Classification/Stroke Lesion Segmentation §Stroke Lesion Generative Network §Demo- CAIDE m: Studio BSR

Generative Adversarial Networks (GANs)

• Two networks competing against each other in a zero sum game

Source Image from https://twitter.com/ch402/status/793911806494261248 Source Image from https://www.slideshare.net/ckmarkohchang/generative- adversarial-networks

(z) (x) The discriminator (D): Distinguish real data from fake created by the generator The generator (G): Learn distribution of the data from random noise, in an attempt to fool the discriminator

Image to Image Translation

• for Generating Stoke Lesion Images

Source image from, Phillip, et al. "Image-to-image translation with conditional adversarial networks." arXiv preprint (2017)

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)

Stoke Lesion Labels Lesion Generated Image Target (Stroke Lesion) GAN for Generating Stoke Lesion Images

Training Pix2Pix-Tensorflow

l Trained conditional GAN below conditions q

Total data set : 2,647 patients (corresponding 33,391 well labeled images) : 80% training, 20% testing

q

Learning parameters: Learning rate =0.0002, L1 weight=100, and GAN weight=1.0

q

About 16 hour up to 200 epoch on NVIDIA Tesla V100 (1 GPU)

SDH SAH SAH

Input Target Output Input Target Output

Examples of Generated Fake CT Image after 200 Epoch Training

EDH SAH

SAH, EDH SAH IVH, IPH SAH, IVH SAH, IVH, IPH

Input Target Output Input Target Output

IPH, SAH SAH, IVH

Evaluation

l In general, evaluating GANs is difficult

q

Loss function makes it harder during training

q

FCN /Inception scores and Amazon Mechanical Turk (AMT by human)

l FCN scores : fake/generated images inferred by FCN l Clarity- threshold blurriness (variance of Laplacian) l Second discriminator- choose more realistic images

Dice: .981 Dice: .976 Dice: .981 Dice: .978

10 epochs 50 epochs 100 epochs 200 epochs

IPH FCN scores (DICE) vs Image Quality

Recall DICE (FCN Scores)

Evaluation – Recall, DICE

l Evaluated by varying:

q

Percentage of training data (based on patient number): 2.5, 10, 50, and 100%

q

Number of epochs : 10, 50, 100, and 200 epoch

Evaluation- Data Augmentation

Original (Real Data) Original+ Augment Original+ 2xAugment Original+ 3xAugment

CAIDE m: Studio BSR (Demo), Booth #726