Abstract Background Magnetic resonance spectroscopy (MRS) complements conventional MRI in brain tumor diagnosis by measuring metabolites associated with various molecular pathways. However, inconsistent and inaccurate voxel placement, due to limited experience and time constraints, has hindered its clinical use. There is a critical need for greater precision and efficiency in MRS voxel placement, which could be addressed by an algorithm leveraging tumor anatomy to facilitate the automation of this process. This pilot study aimed to design and validate an automated, AI-driven MRS voxel placement tool to evaluate suspected diffuse gliomas. Methods In this pilot study, the MRS single voxel auto-placement algorithm utilizes tumor sub-compartment segmentation from a pre-trained deep learning model. Preprocessing involves co-registration and isotropic resampling of multiparametric MRI. The algorithm iteratively adjusts the position and rotation of an MRS voxel mask over the segmented images to maximize enhancing tumor core inclusion while excluding necrosis. Two board-certified neuroradiologists independently and blindly assessed algorithm- and clinically-placed MRS voxels (n = 14) using a Likert scale. Results Fourteen cases of MRS single-voxel placement were evaluated (median age and interquartile range, 47 years 24–64 years; male, 10 patients 71%). Overall quality ratings of 4 or 5 were assigned to 75% of clinically-placed voxels and 79% of algorithm-placed voxels, without any statistical difference (P = .56). Similarly, no statistically significant differences were observed in position ratings (P = .30) or rotation ratings (P = .51). Conclusions No statistically significant differences were observed between algorithm- and clinically-placed voxels, highlighting the promise of AI-based automated MRS single-voxel placement.
Chadha et al. (Fri,) studied this question.