Lower-extremity ulcers are common, with an estimated prevalence is as high as 1-2% among U.S. adults.

The disease burdens are high: 25% to 50% of leg ulcers and more than 30% of foot ulcers are not fully healed after 6 months of treatments, even with the best available ulcer car

The ulcer cares in U.S. are costly. The annual costs associated with venous leg ulcers are approximately $14.9 billions. The annual costs associated with diabetic foot ulcers are approximately $ 9 billions to $ 13 billions. Currently there are 4 major lower-extremity ulcer types: venous ulcers, arterial ulcers, diabetic ulcers and pressure ulcers. Since there are many risk factors and cause, to recognize and classify lower-extremity ulcers are important in diagnosis, treatment and management of risk factors. With the advent of telemedicine, images of lower-extremity ulcers are more than before. There are potentials and needs to use deep learning to recognize and learn appearances, location and patho-physiological features, so AI tools can assist ulcer cares.

In my proposal of image studies lower-extremity ulcers with deep learning, l will use Convolutional Neural Networks (CNNs) to improve the accuracy of image classification.

Convolutional Neural Network (CNN) is special type Neural Networks that works in the same way of a regular neural network except that it has a convolution layer at the beginning. Instead of feeding the entire image as an array of numbers, the image is broken up into a number of tiles, the machine then tries to predict what each tile is. Finally, the computer tries to predict what’s in the picture based on the prediction of all the tiles. This allows the computer to parallelize the operations and detect the object regardless of where it is located in the image.

Convolutional Neural Network (CNNs) can achieve reasonable performance on hard visual recognition tasks — matching or exceeding human performance in some domains. There are successive models continue to show improvements, each time achieving a new state-of-the-art result: QuocNet, AlexNet, Inception (GoogLeNet), BN-Inception-v2, and Inception-v3. For example, if Inceptio-v3 is used, first we could learn how to classify images into 1000 classes in Python or C++. Then we can learn how to extract higher level features from this model which may be reused for other vision tasks.

Training dataset:

Since images of lower-extremity ulcers usually need pre-processing, may not mutually exclusive, could be larger, may not easily labelled, and with variable degrees of noise,
I decide to use dataset with some classes that are mutually exclusive, clearly labelled, smaller size,and without noise, such as CIFAR-10 dataset with considerably much less pre-processing.

Building Ulcer dataset:

I propose to collect large quantities of different lower-extremity ulcers image. Initially, I will start with images with pre-processing to make images more generic, labelled as clearly as possible, reduce noise as I can.
I will then split the dataset into a number of batches for CNNs to learn.
I will gradually advance with overlapping appearances, and locations, increase the levels of difficulty for CNNs to recognize and to learn lower-extremity ulcers, and and finally build and train the model of lower-extremity ulcers. Eventually I will validate dataset and model of lower-extremity ulcers by human ulcer experts.


Author: C. Charles Lin

Status: Project Concept