How much anatomy do we need ? Automated vs . manual pattern recognition of 3 D 1 H MRSI data of patients with prostate cancer

Objective To evaluate quantitatively 3D proton MR spectroscopic imaging (3D H MRSI) data of the prostate in patients with known prostate cancer using anatomical knowledge, compared with a single–voxel–spectra evaluation by blinded experts and automated processing methods. Material and Methods: MRSI data of 10 patients with histologically proven prostate adenocarcinoma, scheduled either for prostatectomy or intensity–modulated radiation therapy, were evaluated by two MRS experts using information on anatomy and localization of the spectra; and in blinded, randomized lists. Spectra were also classified by automated processing methods based on spectral fitting and pattern recognition. Results were compared using Kendall’s tau and grouped in a hierarchical segmentation. Results: The experts came to more binary decisions – using information of spectra from surrounding tissue in ambiguous cases – when evaluating MR spectroscopic images as a whole. Differences between unblinded and blinded evaluation were larger than differences between blinded and automated processing methods. Conclusion: An automated approach can be as good as a blinded reader. However, anatomical and morphological information are routinely used by human experts, supporting their final decisions. Using automated approaches considering anatomical knowledge could therefore improve an automated evaluation, supporting the human reader.