CONVOLUTIONAL DEEP LEARNING
NEURAL NETWORK FOR RADIOLOGICAL
CHEST X-RAY DIAGNOSIS
ISDS 577 Capstone Seminar
Guided by Professor Daniel Soper
Team Members: Skandha Chinta,
Scott Cunningham, Apurva Desai,
Saket Dhamne
CAN A CONVOLUTIONAL NEURAL NETWORK BE USED
TO DETERMINE IF A PATIENT X-RAY HAS A
DIAGNOSABLE CONDITION?
One Tailed T-Test:
Is the mean grayscale value of an X-Ray with
a condition present > than the mean gray
scale value of a X-Ray with no condition?
Image Classification (Deep Learning CNN):
Identify whether an X-Ray has a diagnosable
condition with reasonable accuracy.
Research on identifying specific types of cancer
DATASET FROM NATIONAL INSTITUTE
OF HEALTH
108,948 X-Ray images
from 32,717 unique
patients from one hospital.
01
14 disease categories
image findings and
normal.
02
Made publicly available in
September 2017 for
research purposes.
03
NEURAL NETWORKS
o Assigns weights to input values and
runs through a ‘black box’ of
interconnected data points to
determine which nodes should receive
greatest weights for a given output set.
o Several types of Neural Networks for
different applications.
o Attributes of Convolutional Neural
Networks
• Difficult to train
• Very good at image classification
GOOGLE COLABORATORY
Google’s online testing space
(much like Jupyter)
Uses Gigabyte processors
free
Vastly reduces processing time with
millions of calculations
Python lab
DATA PRE-PROCESSING
Process Reason
Width x Height matching Consistency,
Aspect Ratio is off
AP Labeling AP is preferred over PA
Unique Patients To avoid bias
Latest patient images To avoid bias
Unusual and inaccurate
X-Rays
Eg: You guessed it right, a
PELVIS!
Data Reduction
PRE-
PROCESSING
: T-TEST
• Need to prepare data for T-Test:
• Transform images to mean
grayscale values
• Per image and per cancer type
• Separate into cancer types
• Check outliers, skewness, kurtosis
• Use Histograms, Boxplots and IQR
CHECK ON OUTLIERS, KURTOSIS, SKEWNESS
No_Findings
Quartile 1 112.5463085
Quartile 3 130.4091029
Inter-Quartile Range 17.8627944
Upper bound 157.2032945
Lower bound 85.75211692
Findings
Quartile 1 111.5376172
Quartile 3 128.6265164
Inter-Quartile Range 17.08889915
Upper bound 154.2598651
Lower bound 85.90426848
CHECK ON OUTLIERS, KURTOSIS, SKEWNESS
PRE-PROCESSING: NEURAL NETWORK
Convert all images from
.png to .jpg
Photoshop image filters to
emphasize disease
characteristics
• Crop, Median filter, Smart
Sharpen (edges)
• Exported images in only 2
colors (black & white)
Separated images into
Training (70%) and Testing
(30%)
• Within each divided by
Findings & No Findings
• Compatibility with
neural network
code
T-TEST ON THE BASIS OF VISUAL
PERCEPTION
Grayscale
Normal
Lung
Diseased
Lung
F-TEST TO PERFORM
THE CORRECT T-TEST
• Comparison of
Variance
• If F-Test Value ~= F
Critical Value
We choose T-Test for
two samples assuming
equal variances.
T-TEST
HYPOTHESIS:
• Null Hypothesis based on
Descriptive Statistics
H0 : μgNF >= μgF
• Alternate Hypothesis based on
Visual Perception
HA : μgNF < μgF
T-TEST RESULTS
• Reject the Null Hypothesis with
1% significance level.
• Mean grayscale value of diseased Lung X-
ray
is greater than that of a normal Lung X-
Ray.
• Although T-Test is a statistical test
different from image-classification
techniques, it can be considered a basis
for visual perception.
NEURAL NETWORK
Neural Network architecture for image processing
using Keras library
STEP 1: BUILDING A
CONVOLUTIONAL
LAYER
Number of
Convolutional
Layers - 5
01
Conv2D Function:
• Number of Input filters
varies with the
Convolutional Layer
• Kernel Size
• Activation function =
‘RELU’ (Rectified Linear
Unit)
02
STEP 2: POOLING
• Pooling layer is used to aggregate
information within a small region
of input features and then down
sampling the results
Max Pooling, Pool Size
Strides
Padding
Dropout Layer
STEP 3:
FLATTENING
• What is Flattening?
• Purpose of
Flattening
STEP 4: FULL
CONNECTION LAYER
& OUTPUT LAYER
Purpose to create a fully connected layer
Fully Connected Layer is Hidden Layer
Number of nodes in Hidden Layer
Activation Function for Hidden Layer: RELU
Activation Function for Output Layer: Sigmoid
NEURAL NETWORKS:
OTHER HYPER
PARAMETERS
“Adam” Optimizer
“Binary” Cross Entropy
Evaluation Metrics: Accuracy & Loss
Batch Size: 32
Epochs: 20
Steps per Epochs: 100
METRICS
RESULTS
METRICS: GRAPHICAL REPRESENTATION
RESULTS OF CONVOLUTIONAL
NETWORK MODEL
Model Name Validation Loss Validation
Accuracy
Convolutional Network
Model (Version 55)
0.66 63.13%
PRE-TRAINED MODELS
• Different types of available pre-trained models:
• ResNet50 Model
• MobileNet Model
• Comparison of Different Models:
Model Name Validation Loss Validation Accuracy
Convolutional Network Model
(Version 55)
0.66 63.13%
ResNet50 Model 7.85 51.32%
MobileNet Model 6.82 47.50%
Neuron activation is focused by the darker pixels
HEATMAP FOR TRAINED MODEL
CONCLUSION: MODEL EVALUATION
• Validation Loss
• As close to 0 as possible
• Validation Accuracy
• As a percentage
• CNN Model outperformed pre-trained models: ResNet &
MobileNet
• Best image classification accuracy until 2012 was ~73.8%
FUTURE RESEARCH
• With continued tuning of Hyperparameters and image pre-
processing radiological tools can be developed to identify and flag
abnormal X-Rays for medical attention
• For Future research, we want to train the model to classify by cancer
type.
• Using heatmaps to identify neuron activation for different lesion
types.
THANK YOU

Project presentation - Capstone

  • 1.
    CONVOLUTIONAL DEEP LEARNING NEURALNETWORK FOR RADIOLOGICAL CHEST X-RAY DIAGNOSIS ISDS 577 Capstone Seminar Guided by Professor Daniel Soper Team Members: Skandha Chinta, Scott Cunningham, Apurva Desai, Saket Dhamne
  • 2.
    CAN A CONVOLUTIONALNEURAL NETWORK BE USED TO DETERMINE IF A PATIENT X-RAY HAS A DIAGNOSABLE CONDITION? One Tailed T-Test: Is the mean grayscale value of an X-Ray with a condition present > than the mean gray scale value of a X-Ray with no condition? Image Classification (Deep Learning CNN): Identify whether an X-Ray has a diagnosable condition with reasonable accuracy. Research on identifying specific types of cancer
  • 3.
    DATASET FROM NATIONALINSTITUTE OF HEALTH 108,948 X-Ray images from 32,717 unique patients from one hospital. 01 14 disease categories image findings and normal. 02 Made publicly available in September 2017 for research purposes. 03
  • 4.
    NEURAL NETWORKS o Assignsweights to input values and runs through a ‘black box’ of interconnected data points to determine which nodes should receive greatest weights for a given output set. o Several types of Neural Networks for different applications. o Attributes of Convolutional Neural Networks • Difficult to train • Very good at image classification
  • 5.
    GOOGLE COLABORATORY Google’s onlinetesting space (much like Jupyter) Uses Gigabyte processors free Vastly reduces processing time with millions of calculations Python lab
  • 6.
    DATA PRE-PROCESSING Process Reason Widthx Height matching Consistency, Aspect Ratio is off AP Labeling AP is preferred over PA Unique Patients To avoid bias Latest patient images To avoid bias Unusual and inaccurate X-Rays Eg: You guessed it right, a PELVIS! Data Reduction
  • 7.
    PRE- PROCESSING : T-TEST • Needto prepare data for T-Test: • Transform images to mean grayscale values • Per image and per cancer type • Separate into cancer types • Check outliers, skewness, kurtosis • Use Histograms, Boxplots and IQR
  • 8.
    CHECK ON OUTLIERS,KURTOSIS, SKEWNESS No_Findings Quartile 1 112.5463085 Quartile 3 130.4091029 Inter-Quartile Range 17.8627944 Upper bound 157.2032945 Lower bound 85.75211692 Findings Quartile 1 111.5376172 Quartile 3 128.6265164 Inter-Quartile Range 17.08889915 Upper bound 154.2598651 Lower bound 85.90426848
  • 9.
    CHECK ON OUTLIERS,KURTOSIS, SKEWNESS
  • 10.
    PRE-PROCESSING: NEURAL NETWORK Convertall images from .png to .jpg Photoshop image filters to emphasize disease characteristics • Crop, Median filter, Smart Sharpen (edges) • Exported images in only 2 colors (black & white) Separated images into Training (70%) and Testing (30%) • Within each divided by Findings & No Findings • Compatibility with neural network code
  • 11.
    T-TEST ON THEBASIS OF VISUAL PERCEPTION Grayscale Normal Lung Diseased Lung
  • 12.
    F-TEST TO PERFORM THECORRECT T-TEST • Comparison of Variance • If F-Test Value ~= F Critical Value We choose T-Test for two samples assuming equal variances.
  • 13.
    T-TEST HYPOTHESIS: • Null Hypothesisbased on Descriptive Statistics H0 : μgNF >= μgF • Alternate Hypothesis based on Visual Perception HA : μgNF < μgF
  • 14.
    T-TEST RESULTS • Rejectthe Null Hypothesis with 1% significance level. • Mean grayscale value of diseased Lung X- ray is greater than that of a normal Lung X- Ray. • Although T-Test is a statistical test different from image-classification techniques, it can be considered a basis for visual perception.
  • 15.
    NEURAL NETWORK Neural Networkarchitecture for image processing using Keras library
  • 16.
    STEP 1: BUILDINGA CONVOLUTIONAL LAYER Number of Convolutional Layers - 5 01 Conv2D Function: • Number of Input filters varies with the Convolutional Layer • Kernel Size • Activation function = ‘RELU’ (Rectified Linear Unit) 02
  • 17.
    STEP 2: POOLING •Pooling layer is used to aggregate information within a small region of input features and then down sampling the results Max Pooling, Pool Size Strides Padding Dropout Layer
  • 18.
    STEP 3: FLATTENING • Whatis Flattening? • Purpose of Flattening
  • 19.
    STEP 4: FULL CONNECTIONLAYER & OUTPUT LAYER Purpose to create a fully connected layer Fully Connected Layer is Hidden Layer Number of nodes in Hidden Layer Activation Function for Hidden Layer: RELU Activation Function for Output Layer: Sigmoid
  • 20.
    NEURAL NETWORKS: OTHER HYPER PARAMETERS “Adam”Optimizer “Binary” Cross Entropy Evaluation Metrics: Accuracy & Loss Batch Size: 32 Epochs: 20 Steps per Epochs: 100
  • 21.
  • 22.
  • 23.
    RESULTS OF CONVOLUTIONAL NETWORKMODEL Model Name Validation Loss Validation Accuracy Convolutional Network Model (Version 55) 0.66 63.13%
  • 24.
    PRE-TRAINED MODELS • Differenttypes of available pre-trained models: • ResNet50 Model • MobileNet Model • Comparison of Different Models: Model Name Validation Loss Validation Accuracy Convolutional Network Model (Version 55) 0.66 63.13% ResNet50 Model 7.85 51.32% MobileNet Model 6.82 47.50%
  • 25.
    Neuron activation isfocused by the darker pixels HEATMAP FOR TRAINED MODEL
  • 26.
    CONCLUSION: MODEL EVALUATION •Validation Loss • As close to 0 as possible • Validation Accuracy • As a percentage • CNN Model outperformed pre-trained models: ResNet & MobileNet • Best image classification accuracy until 2012 was ~73.8%
  • 27.
    FUTURE RESEARCH • Withcontinued tuning of Hyperparameters and image pre- processing radiological tools can be developed to identify and flag abnormal X-Rays for medical attention • For Future research, we want to train the model to classify by cancer type. • Using heatmaps to identify neuron activation for different lesion types.
  • 28.

Editor's Notes

  • #3 Speak about the problem statement of the project.
  • #27 Mentioning the accuracy was for a different dataset Never-ending process Metrics Decreased the validation loss from around 8 to close to 0.66 Increased the validation accuracy from approximately 51% to 64% by changing the hyperparameters