Segmentation of adipose and glandular tissue for breast tomosynthesis imaging using a 3D hidden-Markov model trained on breast MRIs

Breast tomosynthesis involves a restricted number of images acquired in an arc in conventional mammography projection geometry. Despite its angular undersampling, tomosynthesis projections are reconstructed into a volume at a dose comparable to mammography. Tomosynthesis thus provides depth information, which is especially beneficial to patients with dense breasts. Because the device can be based on an existing FFDM unit, tomosynthesis may be used to accurately assess breast tissue composition, which would greatly benefit high-risk patients with less access to costly imaging modalities such as MRI. This study plans to extract quantitative 3D breast tissue density information using a fully automatic probabilistic model trained on segmented MRIs. The MRI ground truth was obtained for 293 breasts by iterative threshold-based fatty / glandular tissue segmentation. After training a 3D hidden Markov model (HMM) on 10 MR volumes, our model was validated by segmenting 214 of the 293 breasts. After the tomosynthesis value optimization, the same trained HMM was tested to segment breast tomosynthesis volumes of subjects whose MRIs were used for validation. Initial training / testing of the HMM on MRIs matched density to thresholding within 5% for 70/214 breasts and 10% for 127/214 breasts. HMM segmentation was qualitatively superior at the cranial/caudal end slices in MRIs and quantitatively superior for most tested tomosynthesis volumes. Its robustness and ease of modification give the HMM great promise and potential for expansion in this multi-modality study. © 2011 SPIE.

Duke Scholars

Author Joseph Yuan-Chieh Lo Radiology