Automatic identification of atypical clinical fMRI results

Abstract

Purpose: Functional MRI is not routinely used for neurosurgical planning despite potential important advantages, due to difficulty of determining quality. We introduce a novel method for objective evaluation of fMRI scan quality, based on activation maps. A template matching analysis (TMA) is presented and tested on data from two clinical fMRI protocols, performed by healthy controls in seven clinical centers. Preliminary clinical utility is tested with data from low-grade glioma patients.

Methods: Data were collected from 42 healthy subjects from seven centers, with standardized finger tapping (FT) and verb generation (VG) tasks. Copies of these "typical" data were deliberately analyzed incorrectly to assess feasibility of identifying them as "atypical." Analyses of the VG task administered to 32 tumor patients assessed sensitivity of the TMA method to anatomical abnormalities.

Results: TMA identified all atypical activity maps for both tasks, at the cost of incorrectly classifying 3.6 (VG)-6.5% (FT) of typical maps as atypical. For patients, the average TMA was significantly higher than atypical healthy scans, despite localized anatomical abnormalities caused by a tumor.

Conclusion: This study supports feasibility of TMA for objective identification of atypical activation patterns for motor and verb generation fMRI protocols. TMA can facilitate the use and evaluation of clinical fMRI in hospital settings that have limited access to fMRI experts. In a clinical setting, this method could be applied to automatically flag fMRI scans showing atypical activation patterns for further investigation to determine whether atypicality is caused by poor scan data quality or abnormal functional topography.

Keywords: Brain function; Clinical fMRI; Functional MRI; Language; Motor cortex.

Fig. 1

Schematic overview of the application of automatic assessment of typicality for clinical fMRI

Fig. 2

Rendered group activation patterns per site and per task protocol (|t| < 3.14; df = 6; positive activity in red; negative activity in blue; FT, finger tap protocol; VG, verb generation protocol; L, left; R, right)

Fig. 3

TMA results for each subject per site (a FT (finger tapping), b VG (verb generation)). Horizontal bars represent mean TMA per site. While all centers showed high TMA values, we did find a significant difference between centers in the mean values (F(6,35) = 3.17; p = 0.009), which is likely the result of slight differences in signal to noise between the scanners of the different centers

Fig. 4

Distribution of template matching analysis (TMA) values for typical scans (“TYPICAL,” blue), as well as for atypical scans due to reversed orientation (“REVERSED,” green), incorrect task protocol (“TASK,” orange), or incorrect analysis with a time-shifted regressor (“SHIFT,” yellow). a Finger tapping (FT) template. b Verb generation (VG) template

Fig. 5

Receiver operating characteristic curve for healthy subjects data of a FT and b VG tasks. Figures display the percent correct classification (y-axis), for a range of TMA values

Fig. 6

Distribution of template matching analysis (TMA) values for the verb generation (VG) template for correctly analyzed healthy controls (“TYPICAL,” blue), tumor patient data (“PAT,” red), as well as for atypical scans due to reversed orientation (“REVERSED,” green), incorrect task protocol (“TASK,” orange), or incorrect analysis with a time-shifted regressor (“SHIFT,” yellow). Patient TMA values were significantly higher than SHIFT (t(72) = 8.33; p < 0.001), TASK ((t(72) = 8.33; p < 0.001), as well as REVERSED scans ((t(72) = 8.33; p < 0.001)