A particle filter based autocontouring algorithm for lung tumor tracking using dynamic magnetic resonance imaging

This study introduces a novel autocontouring algorithm based on particle filter for lung tumors. It is validated on dynamic magnetic resonance (MR) images and is developed in the context of MR-linac treatments.

A sequential Monte Carlo method called particle filter is used as the main structure of the algorithm and is combined with Otsu's thresholding technique to contour lung tumors on dynamic MR images. Four non-small cell lung cancer (NSCLC) patients were imaged with a 1.5 T MR for 60 s at a rate of 4 images/s and were asked to breathe normally. Prior to treatment, some image processing is required by the proposed algorithm, which includes a manual contour of the tumor, the tumor's displacement, and its descriptive statistics. During treatment, the contours are automatically generated by thresholding around the center of mass of the particles. A comparison with the expert's contours is obtained by calculating the Dice similarity coefficient (DSC), the precision, the recall, the Hausdorff distance, and the difference in centroid positions (Δd).

This autocontouring algorithm is independent of pretreatment training and presents continuous adaptability as provided by the nature of particle filters. The number of particles is proportional to the area of the tumor and increases the computational time at a rate of 2 ms for every 500 particles, whereas the contouring step adds a constant 14 ms. The contours' comparison is obtained with a mean DSC of 0.89-0.91, mean precision of 0.88-0.91, mean recall of 0.89-0.95, and mean Δd of 0.6-2.0 mm.

This work presents a proof of concept of a new autocontouring algorithm for NSCLC patients on dynamic MR images. The contours were generated in good agreement with the expert's contours.