Convolutional Neural Networks for Medical Images Diagnosis

Name: Convolutional Neural Networks for Medical Images Diagnosis
Rating: 4.2 (202 reviews)

CNN, Deep Learning, Medical Imaging, Transfer Learning, CNN Visualization, VGG, ResNet, Inception, Python & Keras

Created byHussein Samma

Last updated 6/2020

English

What you'll learn

To build from scratch a CNN-based medical diagnosis model.
To learn how to get and prepare medical dataset used in this work.
To understand by examples how CNN layers are working.
To learn by examples different measures which used to evaluate CNN.
To learn different techniques used to improve the performances of CNN.
To learn how to visualize CNN intermediate layers.
To learn how to deploy the trained CNN model using flask API server.
To learn how to implement all steps using python, tensorflow, and keras.

Course content

9 sections • 29 lectures • 1h 29m total length

Introduction0:42
Build and deploy a CNN-based medical diagnosis model from scratch for remote, web-based access. Explore practical techniques in computer vision, deep learning, and pattern recognition for medical image diagnosis.
Outlines3:54

Overview of CNN Architecture5:08
Explore the CNN architecture, contrasting end-to-end deep learning with traditional hand-crafted features, and learn how convolution, ReLU, and max-pooling extract features for medical image diagnosis.
CNN Convolutional Layer2:49
Explore the convolutional operation in CNNs by applying 3x3 filters with zero padding to a 6x6 input. Learn how to implement this in Python with a 32-filter convolutional layer.
CNN ReLU Activation1:13
Apply the ReLU activation to the convolution output, turning negatives to zero and keeping positives, with a Python implementation and placement as a new layer before max-pooling.
CNN Max-pooling Layer1:17
Learn how max-pooling follows ReLU, reduces data by applying a 2-by-2 window and taking the maximum from each region, and how to implement this layer in Python.
CNN Flattening Layer1:07
CNN Fully Connected Layer1:51
Explore the fully connected layer after the flattening step in CNNs, where inputs connect to neurons through weights and bias, and ReLU activation is applied.
CNN Softmax Classifier2:25
Demonstrates the softmax classifier as the CNN's last layer, using two fully connected neurons for normal and abnormal classes, computing probabilities from inputs, weights, and biases.

Data Normalization3:02
Master data normalization before training a CNN for medical image diagnosis. Normalize grayscale pixel values by scaling to 0-1, shifting to zero mean, and standardizing to unit variance using Keras.
Cross Entropy Loss function1:07
Learn how cross entropy loss function guides CNN training for medical image diagnosis by computing error from predicted class probabilities against true labels.
CNN Backpropagation3:22
Python Implementation of CNN Training11:21
Implement python cnn training for chest x-ray classification, organize data into training and validation sets, build a 3-block cnn with 32–64–128 filters, train with sgd and cross-entropy, and save model.

CNN Hyper-Parameters Tuning2:42
Analyse the effect of pooling layer.
Using Different Optimizers2:03
Data Augmentation1:25
Explore how data augmentation using rotation, shifting, zooming, and brightness via the image data generator improves CNN performance, reduces overfitting, and raises ROC AUC to 0.91 in medical image diagnosis.
Using Dropout Regularization1:49
Batch Normalization2:10
Leaky ReLU Activation1:54
Replace the ReLU with LeakyReLU in a CNN to address zero gradients, improving performance on medical image diagnosis and achieving a roc area of 0.92.
Early Stopping1:30
Explains early stopping in cnn training by monitoring validation accuracy and halting if there is no improvement for 10 epochs, reducing training to 30 epochs while preserving testing accuracy.

Transfer Learning using Oxford VGG Network5:07
Transfer learning with shallow VGG network
Transfer Learning using Microsoft ResNet Network2:44
Explore transfer learning with the Microsoft ResNet network, highlighting deep residual architecture with branches and shortcuts, replaced fully connected layers by average pooling. Results show limited gains versus transfer VGG.
Transfer Learning using Google Inception Network2:20
Learn transfer learning with the Google Inception network for medical image diagnosis, comparing Inception v3 and v4, and replacing the classification layer with an average pooling classifier.

Requirements

Have the basic knowledge about CNN
Familiar with Python programming
Spyder editor with Python 3.7

Description

This course was designed and prepared to be a practical CNN-based medical diagnosis application. It focuses on understanding by examples how CNN layers are working, how to train and evaluate CNN, how to improve CNN performances, how to visualize CNN layers, and how to deploy the final trained CNN model.

All the development tools and materials required for this course are FREE. Besides that, all implemented Python codes are attached with this course.

Who this course is for:

This course was designed for students who are interested in the applications of CNN to solve real-world medical diagnosis problem.

Convolutional Neural Networks for Medical Images Diagnosis

What you'll learn

Explore related topics

Course content

Introduction2 lectures • 5min

Getting and Preparing Data1 lecture • 3min

CNN Architecture7 lectures • 16min

CNN Training4 lectures • 19min

CNN Evaluation3 lectures • 15min

CNN Improvement7 lectures • 14min

CNN Transfer Learning3 lectures • 10min

CNN Visualization1 lecture • 3min

CNN Deployment1 lecture • 5min

Requirements

Description

Who this course is for: