An algorithm for longitudinal registration of PET/CT images acquired during neoadjuvant chemotherapy in breast cancer: preliminary results

Abstract

Background: By providing estimates of tumor glucose metabolism, 18F-fluorodeoxyglucose positron emission tomography (FDG-PET) can potentially characterize the response of breast tumors to treatment. To assess therapy response, serial measurements of FDG-PET parameters (derived from static and/or dynamic images) can be obtained at different time points during the course of treatment. However, most studies track the changes in average parameter values obtained from the whole tumor, thereby discarding all spatial information manifested in tumor heterogeneity. Here, we propose a method whereby serially acquired FDG-PET breast data sets can be spatially co-registered to enable the spatial comparison of parameter maps at the voxel level.

Methods: The goal is to optimally register normal tissues while simultaneously preventing tumor distortion. In order to accomplish this, we constructed a PET support device to enable PET/CT imaging of the breasts of ten patients in the prone position and applied a mutual information-based rigid body registration followed by a non-rigid registration. The non-rigid registration algorithm extended the adaptive bases algorithm (ABA) by incorporating a tumor volume-preserving constraint, which computed the Jacobian determinant over the tumor regions as outlined on the PET/CT images, into the cost function. We tested this approach on ten breast cancer patients undergoing neoadjuvant chemotherapy.

Results: By both qualitative and quantitative evaluation, our constrained algorithm yielded significantly less tumor distortion than the unconstrained algorithm: considering the tumor volume determined from standard uptake value maps, the post-registration median tumor volume changes, and the 25th and 75th quantiles were 3.42% (0%, 13.39%) and 16.93% (9.21%, 49.93%) for the constrained and unconstrained algorithms, respectively (p = 0.002), while the bending energy (a measure of the smoothness of the deformation) was 0.0015 (0.0005, 0.012) and 0.017 (0.005, 0.044), respectively (p = 0.005).

Conclusion: The results indicate that the constrained ABA algorithm can accurately align prone breast FDG-PET images acquired at different time points while keeping the tumor from being substantially compressed or distorted.

Trial registration: NCT00474604.

Figures

Figure 1

Our own support for prone breast PET/CT images. It allows for the breasts to lie pendant during the scanning procedure, therefore greatly enhancing the ability to perform longitudinal registration. Typical breast PET/CT is performed in the supine position which results in less reproducible patient positioning between scan sessions.

Figure 2

The scheme for applying the algorithm to register the PET/CT. The CT data obtained during t1 and t2 are aligned via the proposed registration algorithm (steps A and B) to the CT images acquired at t3. The resulting DF are then applied to the corresponding PET images to yield co-registered longitudinal PET data (steps C and D).

Figure 3

A representative patient displaying the results of the three registration algorithms. The first three rows correspond to the three time points, and the three columns show the results after rigid body registration, after unconstrained ABA registration (ABA), and with constrained ABA registration (ABA_CON), respectively. In the fourth row, the first panel displays the deformation field generated by the ABA when the images at t1 are registered to the images at t3, while the second panel shows the result using the ABA_CON; the third and fourth panels display similar data when the images at t2 are registered to the images at t3, respectively. The green circle shows the tumor location. The contour of the CT image at t3 is drawn and then copied to other images to facilitate the comparison.

Figure 4

Results of the three registration algorithms for another patient with similar setup with Figure3.

Figure 5

Ten examples from ten patients illustrating the matching accuracy of the registration algorithm. Axial CT slices obtained at t1 or t2 (colored in blue) are overlaid on the corresponding images obtained at t3 (gray) in a checkerboard pattern to facilitate assessments of the registration performance. Note that the structural boundaries between blue and gray images are connected accurately, indicating a good performance of the registration algorithm.