Data-driven breast decompression and lesion mapping from digital breast tomosynthesis.

Digital breast tomosynthesis (DBT) emerges as a new 3D modality for breast cancer screening and diagnosis. Like in conventional 2D mammography the breast is scanned in a compressed state. For orientation during surgical planning, e.g., during presurgical ultrasound-guided anchor-wire marking, as well as for improving communication between radiologists and surgeons it is desirable to estimate an uncompressed model of the acquired breast along with a spatial mapping that allows localizing lesions marked in DBT in the uncompressed model. We therefore propose a method for 3D breast decompression and associated lesion mapping from 3D DBT data. The method is entirely data-driven and employs machine learning methods to predict the shape of the uncompressed breast from a DBT input volume. For this purpose a shape space has been constructed from manually annotated uncompressed breast surfaces and shape parameters are predicted by multiple multi-variate Random Forest regression. By exploiting point correspondences between the compressed and uncompressed breasts, lesions identified in DBT can be mapped to approximately corresponding locations in the uncompressed breast model. To this end, a thin-plate spline mapping is employed. Our method features a novel completely data-driven approach to breast shape prediction that does not necessitate prior knowledge about biomechanical properties and parameters of the breast tissue. Instead, a particular deformation behavior (decompression) is learned from annotated shape pairs, compressed and uncompressed, which are obtained from DBT and magnetic resonance image volumes, respectively. On average, shape prediction takes 26s and achieves a surface distance of 15.80 +/- 4.70 mm. The mean localization error for lesion mapping is 22.48 +/- 8.67 mm.

Optimal classification of long echo time in vivo magnetic resonance spectra in the detection of recurrent brain tumors.

We describe the optimal high-level postprocessing of single-voxel (1)H magnetic resonance spectra and assess the benefits and limitations of automated methods as diagnostic aids in the detection of recurrent brain tumor. In a previous clinical study, 90 long-echo-time single-voxel spectra were obtained from 52 patients and classified during follow-up (30/28/32 normal/non-progressive tumor/tumor). Based on these data, a large number of evaluation strategies, including both standard resonance line quantification and algorithms from pattern recognition and machine learning, were compared in a quantitative evaluation. Results from linear and non-linear feature extraction, including ICA, PCA and wavelet transformations, and also the data from resonance line quantification were combined systematically with different classifiers such as LDA, chemometric methods (PLS, PCR), support vector machines and ensemble methods. Classification accuracy was assessed using a leave-one-out cross-validation scheme and the area under the curve (AUC) of the receiver operator characteristic (ROC). A regularized linear regression on spectra with binned channels reached 91% classification accuracy compared with 83% from quantification. Interpreting the loadings of these regressions, we find that lipid and lactate signals are too unreliable to be used in a simple machine rule. Choline and NAA are the main source of relevant information. Overall, we find that fully automated pattern recognition algorithms perform as well as, or slightly better than, a manually controlled and optimized resonance line quantification.