This work is novel and proposes a different philosophy that departs from more conventional approaches for achieving robust non-invasive prostate tumor analysis in a clinical setting. Studies, such as this work, possess only a limited number of combined multi-parametric MRI (MP-MRI) with verifying concomitant whole mount prostatectomy. The algorithms discussed in this study use multispectral tumor signatures or image-based biomarkers that characterize the tumors. These image-based biomarkers are derived from training sets and digitally stored in a library or repository. Employing image-based biomarkers in this study uses far fewer training sets (tens) versus hundreds or thousands used in other approaches. To accommodate for varying patient sizes, MRI scanners, MRI pulse sequences, MRI calibrations etc., this study instead transforms signatures or image-based biomarkers from a repository or library and then transformed them based on MP-MRI statistics for each patient scan (not in the library) using “Whitening-DeWhitening” transform (Mayer et al. 2002 and 2003). These transformed signatures are inserted into the simple supervised target algorithms using hypercone decision surfaces from Adaptive Cosine Estimator (ACE) and Spectral Angle Mapper (SAM) (Manolakis et al. 2002) rather than hyperplane decision surfaces used to discriminate targets from backgrounds in the case of Linear Discrimination Analysis (LDA) and Support Vector Machines (SVM) or other geometries that use the computationally intense kernel approach. For hyperspectral images, hypercone decisions achieved greater target discrimination than hyperplane decision surfaces (Manolakis et al. 2002). The image-based biomarkers do not employ arbitrary fitting parameters, and they instead simply combine all MRI modalities.
Conventionally, MP-MRI algorithms (Lemaitre et. al. 2015, Niaf et al. 2014, Peng et al. 2013, Qian et al., 2016, Vos et al. 2010 and 2012) use Linear Discrimination Analysis (LDA), Kernels, Random Forests, and/or Support Vector Machines to quantitatively detect and score tumors. These algorithms numerically extract, identify, and associate features with classes derived from training, are computationally intense especially when employing a kernel to linearize the discriminating surface, and use a large number of training sets that presumably span for every potential patient, scanner, etc.. In addition, these large training sets are required for Random Forests (Ho,1998), in order to elevate the number of “trees” and “leaves” in the forest, shrink the variation, and thereby enhance the statistics and classification accuracy.
The purely “spectral” supervised target detection approach discussed in this paper does not use local spatial features such as textures that describe the local roughness, smoothness, entropy etc. (Fehra et al., 2015; Tiwari et al., 2012; Viswanath et al., 2012, Niaf et al., 2014). Spatial feature extraction requires a spatial window spanning several voxels to generate co-occurrence matrices thereby degrading the detection spatial resolution. Some (Kim et al., 2007; Padhani,, 2002; Sato et al., 2005; Vargas et al., 2011) used statistical averages and histograms of a given MRI modality over a given region of interest (ROI) delineated by the radiologist to evaluate the disease over an entire tumor and cannot assess the disease heterogeneity. In contrast, the purely spectral approach described in this article can potentially find and examine much smaller targets, as small as a single voxel, baring adjacency effects from neighboring voxels. Further the vector nature of each voxel is exploited by a color display through judiciously assigning red, green, and blue to three of the seven components. Tumor detection, scoring, and color highlighting relies on the spectral distribution alone and departs from more standard Computer Aided Diagnosis (CAD) algorithms that depend solely on spatial analysis.
Advanced Algorithm and Image-Based Biomarkers for Cancer Detection, Scoring and Tumor Volume Measurements and Color Display