Are You a CNN Expert?

Introduction to Convolutional Neural Networks for Medical Deep Learning

• Convolutional neural networks (CNNs) are the building blocks of most deep learning systems and are used for a broad range of applications, including medical diagnosis, robot vision, and machine translation.

• The basic concept of CNNs involves using convolution filters to extract features from images, which are learned through training data that has been manually labeled with class labels.

• The ImageNet challenge provided 1.2 million RGB images of size 224x224 annotated with one of a thousand classes each, which was difficult to solve using traditional machine learning algorithms until the first deep learning network, VGGNet, was developed.

• In contrast to classical machine learning approaches, CNNs learn features and classifiers jointly in one architecture, without making many handcrafted assumptions about what features should be extracted from the images.

• CNNs achieve locality in convolution filters through sparse connectivity, which connects neighboring pixels in the next layer, and weight sharing, which repeatedly uses the same weights to reduce the number of trainable parameters.

• There are different choices to make when designing a deep learning architecture, such as using dilations, stride, and padding, and applying pooling layers for aggregation and downsampling.

• CNNs extract features from images through a stack of feature or activation maps, which are created by applying a set of filters in one layer, and then inserting a non-linearity and another layer of convolutions with another set of filters.

• The deeper the network, the more abstract features are extracted, combining different object parts into whole objects and full scene and image understanding.

• Pooling layers reduce the spatial resolution and aggregate over a window of 2x2 the maximum activation value, which is not strictly necessary but usually a good idea for downsampling and upsampling.

• Implementing a convolution involves generating a matrix of input from the image and extracting patches from the original image, which is computationally demanding and can be optimized using math libraries and hardware architectures such as GPUs.

• CNNs are used for medical diagnosis, including segmenting the ventricle part of the heart in cardiac MRI scans, and classifying ECG signals, which use 1D convolution instead of 2D convolution.

• CNNs are also used for generating new content, such as realistic video movies of yourself being an experienced dancer by combining your own photos dataset with the video of some experienced dancer, and synthesizing images of faces and medical images without replicating training data.

• In the next lecture, the training data will be employed to find good values for the parameters in these layers using stochastic gradient descent.Understanding Convolutional Networks

• Convolutional networks are used for image classification tasks.

• Convolutional networks consist of multiple layers with pooling in between to reduce the size of the representation.

• Nonlinearities are used after each convolution layer to improve the accuracy of the model.

• Convolutional networks usually end with a fully connected layer that produces a global prediction for the whole image.

• Deep networks have dozens or hundreds of layers, which makes the individual meaning of one feature layer less certain.

• Bag of features is an alternative approach to image classification that uses a histogram of local image features.

• Bag of features is less effective than convolutional networks, but it requires less computation.

• Convolutional networks use filters to extract features from the input image.

• Filters are applied to overlapping or non-overlapping patches of the input image.

• The patches are then put together as a matrix and multiplied with the filter bank to generate output maps.

• The output maps can be put together to form an image again using a fold operation.

• Hardware architectures that are efficient at matrix multiplication are important for the efficient implementation of convolutional networks.