Combining OPM and lesion mapping data for epilepsy surgery planning: a simulation study

Abstract

When planning for epilepsy surgery, multiple potential sites for resection may be identified through anatomical imaging. Magnetoencephalography (MEG) using optically pumped sensors (OP-MEG) is a non-invasive functional neuroimaging technique which could be used to help identify the epileptogenic zone from these candidate regions. Here we test the utility of a-priori information from anatomical imaging for differentiating potential lesion sites with OP-MEG. We investigate a number of scenarios: whether to use rigid or flexible sensor arrays, with or without a-priori source information and with or without source modelling errors. We simulated OP-MEG recordings for 1309 potential lesion sites identified from anatomical images in the Multi-centre Epilepsy Lesion Detection (MELD) project. To localise the simulated data, we used three source inversion schemes: unconstrained, prior source locations at centre of the candidate sites, and prior source locations within a volume around the lesion location. We found that prior knowledge of the candidate lesion zones made the inversion robust to errors in sensor gain, orientation and even location. When the reconstruction was too highly restricted and the source assumptions were inaccurate, the utility of this a-priori information was undermined. Overall, we found that constraining the reconstruction to the region including and around the participant's potential lesion sites provided the best compromise of robustness against modelling or measurement error.

Figure 1

Data simulation and model comparison pipeline. Sensor layouts were created with 95 and 25 OPMs. The lead fields for the Nolte single shell model were calculated with the sensors in their original positions and then with the sensor positions or orientations changed. OP-MEG data at the sensors (in their original positions) were simulated with 4 different source models for each lesion: the centre of mass (COM) of the lesion, the whole lesion active, the whole edge or boundary of the lesion and a piece of the lesion edge. The blue area in panel 2 shows the lesion map for an example lesion, the orange dots are the mesh vertices simulated as being active in each condition. The simulated data was then localised using three different methods: an empirical Bayesian Beamformer (EBB), uninformed of the potential lesion sites, a loosely constrained implementation of MSP, with a patient’s anatomical lesions and the regions around them as priors, and a highly restricted version of MSP with the centre of mass of a patient’s anatomical lesions as priors. For both MSP inversions, each of a patient’s lesions was modelled separately and the Free Energy of the models compared. The lesion/prior with the higher Free Energy was retained.

Figure 2

(A) Joint histogram of age and number of predicted lesions for participants in the MELD dataset. While the modal average number of predicted lesions is one, the median number is two. (B) Distribution of lesion predictions across the cortical surface. There is a higher density of lesions in the temporal or frontal lobes than elsewhere. Based on Ref..

Figure 3

Distribution of the Free Energy difference between the correct (simulated) lesion model and the alternative lesion model with the next highest Free Energy ($\mathrm{\Delta }\mathrm{F}$) for all 1309 lesions from the restricted reconstruction. Values of $\mathrm{\Delta }\mathrm{F}$ below 0 indicate that an incorrect lesion would be selected. The maximum of the distribution and 5th and 95th percentile values are marked with dashed black lines. In (A–D), different sensor errors are shown. Flexible type errors (position (A), orientation (B) and gain uncertainties (C)) have a small effect by comparison with rigid helment rotation (D). Source errors (E), show the largest negative skew in the distribution of $\mathrm{\Delta }\mathrm{F}$. In other words, for the scenarios tested here, incorrect source models undermine the correct lesion choice to a much greater degree than incorrect sensor models.

Figure 4

Comparison of the reconstruction methods tested: unrestricted EBB, MSP loosely constrained around the lesion site and MSP restricted to the centre of mass of the lesion. The percentage of lesions which were correctly differentiated from any of the patient’s other lesions is shown against the different errors added. The highly restricted method performs best when the source model is the same as the simulated source (i.e. the COM of the lesion was used both to simulate and model the data). When the model was not representative of the simulated sources, the loosely constrained method performed the best. In all cases, all methods performed considerably higher than the chance level of 24.4%.

Figure 5

Relationship between the percentage of lesions correctly identified and distance to the nearest lesion from the same patient (i.e. the nearest alternative potential lesion site). All three reconstruction methods are shown for a selection of sensor and source errors. Each datapoint represents the percentage of lesions correctly identified within a distance range (or bin). The distance ranges are sampled so that there are equal numbers of lesions in each distance range (or bin). The points indicate the centre of each bin. The shaded area gives the 95% confidence interval of the percentage correct in each bin, estimated by bootstrapping with the scipy stats toolbox. (A) No errors or uncertainties have been added. (B) Random errors of 5 mm to the OPM positions, 10 degrees to the OPM orientations and 5% to the OPM channel gains have been added. (C,D) 10 degree (C) and 20 degree (D) rigid helmet rotations. (E–G) Sensor errors were set to zero and source errors were introduced, with the whole lesion (E), edge of the lesion (F) and a single vertex on the edge of the lesion (G) simulated while the COM of the lesion was used to model the lesion in the highly restricted reconstruction.