Calculate Breast Density with Artificial Intelligence

Quickly flipping through the pages of the Journal of Digital Imaging (JDI) with eager eyes, I stopped on an article with the title of "Identification and Localization of Endotracheal Tube on Chest Radiographs Using a Cascaded Convolutional Neural Network Approach." After completing my first artificial intelligence (AI) research project, I had finally published an article in a peer-reviewed medical journal to share my findings with the scientific community.

As I read through the article once more, the black-and-white X-ray images in the middle of the page caught my eye, and I remembered my first encounter with AI. When my grandmother was diagnosed with breast cancer, I was determined to find a way to help. I discovered how AI could detect tumors with incredible accuracy by training on thousands of images. I instantly fell in love with the power of machine learning and craved an opportunity to work with it.

After submitting my first article to JDI, I began conducting new research without wasting time. Since my grandmother’s unexpected diagnosis had led me to explore machine learning, I still yearned to find a way for the early diagnosis of breast cancer. As a student research intern at the UCI School of Medicine, I collaborated with other researchers to develop an algorithm that would predict breast density, a strong indicator of breast cancer.

I spent over ten months since September 2020. As an intern at the UCI School of Medicine, I've dedicated at least fifteen hours a week to my research. Therefore, I've worked for more than 600 hours overall on this project.

My previous project used supervised learning, where I had to spend several months manually annotating images. This year, I wanted to investigate unsupervised learning techniques that would let me spend less time annotating images and more time building efficient algorithms. After reading a recently published article about variational autoencoders (VAE), I was determined to create an algorithm to train a VAE model to predict breast density, a strong indicator of breast cancer.

First, I needed to search for a large dataset from the UCI repository. I was able to find a dataset consisting of 734 3D MRIs and 6,987 2D mammograms of 734 unique patients. For each MRI, I identified mammograms of the same patient. The next step was preprocessing. The original breast mammograms in DICOM format were zero-padded and resampled to 512x512.

I calculated breast density as the ratio of fibroglandular tissue (FGT) to the whole breast. Dr. Chang pointed me to a U-Net study for segmentation of the FGT and breast regions on mammograms. I generated the ground-truth values that I could use to correlate with my results using VAE for patch-based training.

I trained the encoder of my VAE on randomly cropped 32x32 patches on 6,987 mammograms to learn contrast. For testing, I ran the encoder model prediction on all 32x32 sections of each 512x512 input image. The encoder collapses each 32x32 patch on the original image down to a single value, like a compression mechanism, leaving me with a 16x16 latent matrix prediction.

Lastly, I used a masking algorithm to exclude the non-breast regions, such as the background and the pectoralis muscle. Once I completed the clean-up procedure, I was ready to generate a number to represent the breast density. I tried several single-value metrics such as the mean, median, 25th percentile, and 75th percentile, and correlated each value with the ground truth from the MRI’s. I tested all hyperparameter settings, so I could compare different configurations of the VAE model and select the combination that generates the best performance. In total, I ran more than 600 model configurations, each one for 20,000 to 40,000 iterations, on 10 GPU servers in parallel.