Evaluation of a deformable registration algorithm for subsequent lung computed tomography imaging during radiochemotherapy

Purpose: Rating both a lung segmentation algorithm and a deformable image registration (DIR) algorithm for subsequent lung computed tomography (CT) images by different evaluation techniques. Furthermore, investigating the relative performance and the correlation of the different evaluation techniques to address their potential value in a clinical setting. Methods: Two to seven subsequent CT images (69 in total) of 15 lung cancer patients were acquired prior, during, and after radiochemotherapy. Automated lung segmentations were compared to manually adapted contours. DIR between the first and all following CT images was performed with a fast algorithm specialized for lung tissue registration, requiring the lung segmentation as input. DIR results were evaluated based on landmark distances, lung contour metrics, and vector field inconsistencies in different subvolumes defined by eroding the lung contour. Correlations between the results from the three methods were evaluated. Results: Automated lung contour segmentation was satisfactory in 18 cases (26%), failed in 6 cases (9%), and required manual correction in 45 cases (66%). Initial and corrected contours had large overlap but showed strong local deviations. Landmark-based DIR evaluation revealed high accuracy compared to CT resolution with an average error of 2.9 mm. Contour metrics of deformed contours were largely satisfactory. The median vector length of inconsistency vector fields was 0.9 mm in the lung volume and slightly smaller for the eroded volumes. There was no clear correlation between the three evaluation approaches. Conclusions: Automatic lung segmentation remains challenging but can assist the manual delineation process. Proven by three techniques, the inspected DIR algorithm delivers reliable results for the lung CT data sets acquired at different time points. Clinical application of DIR demands a fast DIR evaluation to identify unacceptable results, for instance, by combining different automated DIR evaluation methods.