3D characterization of texture: Evaluation for the potential application in mammographic mass diagnosis

Texture is one of the most important features used to characterize and interpret mammographic images within the context of computer assisted diagnosis. As volumetric medical imagery such as digital breast tomosynthesis data, MRI or 3D-US are available for breast diagnosis, we are interested if state of the art 2D texture analysis methods can be extended to 3D, and furthermore, if these extensions can enhance the characterization of depicted lesions. Hence, we want to compare the discrimination ability of the 2D methods with their corresponding 3D extensions. Due to the lack of a sufficient amount of available annotated and histologically validated 3D medical datasets, a computer synthesized volumetric texture database was used to perform the experiments. Three setups were investigated for each texture analysis method: 2D texture on one representative single slice of the 3D-volume, a so-called 3D "intra-slice" approach, concatenating texture information from adjacent 2D-slices, and the so-called "3D inter-slice" approach. A genetic algorithm was used to select an optimal feature subset for each approach and setup. The classification performance of these subsets was evaluated using a SVM classifier. The classification accuracy ranged from 62% to 86% for the 2D approaches, from 66% to 91% for the 3D intra-slice approaches and from 70 % to 85 % for the 3D inter-slice approaches. For every approach one of the 3D setups outperformed the 2D setup. Hence, the extension to 3D enhances the classification accuracy on the synthesized volumetric texture database. In a next step these results have to be validated on real medical image data.