Optimization of Hidden Markov Model for Density Estimation in Breast Imaging

Abstract

Breast cancer is the second leading cause of cancer-related death in women in the United States; it is fundamentally important to detect risks as accurately and as early as possible for prevention. As breast density often correlates with risk factors, 3D Hidden Markov Model method is one quantitative approach to measuring density in breast tomosynthesis images. The purpose of developing this method was to overcome the poor depth resolution in tomosynthesis images and to get a more accurate density estimation of a woman's breast through 3D imaging. Previous study showed that 3D Hidden Markov Model was able to accurately measure breast density for tomosynthesis images that matched closely with the defined ground truth. However qualitatively looking at the segmentation, there were noise and voxels segmented as glandular or adipose that did not reflect the breast anatomy. In addition, the model training process was very time consuming and required many patient cases. To address these limitations, this research was conducted to optimize three of the key model parameters, namely the intensity rescaling of the images, the spatial relation of the neighbors to the current voxel, and the number of training cases. The HMM model was trained using 100 breast MRI volume then validated on 12 breast tomosynthesis volumes.The HMM model was simplified by cutting the number of grayscale intensity levels in half, while sampling distance between neighboring voxels was doubled and tripled. In spite of these changes, however, there was no significant change to the breast density estimation of the tomo images. We identified new challenges in matching grayscale histogram intensities between breast MRI and the two breast tomosynthesis prototype models, and we gained understanding of how the neighbor's distance brought in different segmentation characteristics.