Conventional and multi-state cellular neural networks in segmenting breast region from MR images: Performance comparison

Automated evaluation of MR images for breast density assessment or lesion localization requires accurate segmentation of breast region from regions of the body, such as the chest muscle, lungs, heart and ribs. Breast region segmentation is very complicated in the presence of background noise, intensity inhomogeneity and partial volume artifacts on MR images. Cellular neural networks (CNNs) are massively parallel cellular structures with locally interconnected cells and learning abilities and offer efficient ways to perform many complex medical image segmentation tasks. In this study, the performance of two breast region segmentation methods based on conventional CNNs and multi-state CNNs have been compared using non fat-suppressed T2-weighted bilateral axial images selected from 23 healthy women examined using a 3 Tesla MR scanner. The images provide a range of breast fat content representing 48 fatty, 61 fibroglandular or heterogeneously dense and 28 dense breast slices. Statistical analyses show that multi-state based method performs significantly better with high average precision, high true positive volume fraction, and low false positive volume fraction with an overall performance of 99.3±1.8%, 99.5±1.3%, and 0.1±0.2%, respectively.