Machine Learning Research
Michigan Medicine Radiology Department
MY ROLE
Research
Image Classification
Image Reconstruction
Python Programming
TEAM MEMBERS
Yang Meng
Yanglong Lu
TOOLS USED
Jupyter Notebook
Matlab
TIMELINE
October 2021 - April 2022
Background
Imaging techniques are used in medical applications to diagnose and prognose different diseases. The ability to share clinical data including medical images across different organizations and locations is significant to improve the healthcare system.
Problem
Storage and sharing of image data are challenging because the volume of the image data grows rapidly. Thus, there is a pressing need to develop a framework to share and store medical images efficiently.
Objective
To develop a physics-constrained dictionary learning method to improve the ability to store and manage data that consumes the least amount of space with little to no impact on quality of the image.
Methods
Found a medical database with 4,999 labeled medical images of 14 different diseases. 871 images corresponding to 4 different diseases were pulled out using Python programming from those 14 categories. The collected images are classified into four categories, and one example in each category with the size of 1024Γ1024 pixels is shown below.
To demonstrate the proposed physics-constrained dictionary learning method, the original images with the size of 1024Γ1024 pixels are rescaled to low-resolution ones to the size of 25Γ25 pixels, for classification and reconstruction. Especially in the real-time monitoring process, the amount of data transmitted in communication channels is always limited.
Developed tools for image processing (feature extraction) using Python programming
a. Trained & Tested the labeled images
b. Used the algorithm SVM for image classification
Applied the traditional compressed sensing (CS) technique which means we used a few pixels from the original image to reconstruct the original images and applied dictionary learning for the classification and image reconstruction process using Matlab
Database (Four Different Diseases)
Atelectasis
Effusion
Infiltration
Nodule
Machine Learning
Classification: a process of categorizing a given set of data into categories split the the data set into two sets:
training phase
testing phase (image labeling)
Train/Test: a method to measure the accuracy of your model
SVM: Machine Learning Algorithm used for Classification
Compressed Sensing
The compressed sensing technique is to take a sample of random pixels from our original medical image π, subsample pixels (y) from π and use these subsample pixels to reconstruct the original image efficiently.
y = π½π = π½πΏπΈ = ΞπΈ
Dictionary Learning
Dictionary learning was used in compressed sensing to classify and reconstruct the original image.
Dictionary Learning Method Used:
K-SVD
K-SVD Method Process
(a) Clean image (b) noisy version (c) trained dictionary
(d) corresponding denoised result, using the K-SVD algorithm
Our Approach: Physics-Constrained Dictionary Learning
Physical constraint: the physics-constrained dictionary learning method uses the K-SVD algorithm, but is proposed to solve reconstruction and classification problems
We need the following for the physical-constrained method:
Locations of the collected pixels
Number of pixels collected
Select features/region of interest
Testing
Original Image (that has been shared several times and is 25x25 pixels)
Image I reconstructed using our physics-constrained dictionary learning method
Using our physics-constrained dictionary learning method, we are able to make the image more clearer.
Results
Physics-constrained dictionary learning for classification results:
Fewer pixel values collected = reconstructed images contain more noises
More pixel values collected = reconstructed images can be more blurred
Ultimately, with low-resolution images, high-resolution ones can be reconstructed using our physics-constrained dictionary learning method, which will ultimately improve the efficiency of data storage in medical applications.