A computerized system for segmenting lesions in head and neck CT scans was developed to assist radiologists in
estimation of the response to treatment of malignant lesions. The system performs 3D segmentation based on a level set
model and uses as input an approximate bounding box for the lesion of interest. We investigated the effect of the
interobserver variability of radiologists' marking of the bounding box on the automatic segmentation performance. In
this preliminary study, CT scans from a pre-treatment exam and a post one-cycle chemotherapy exam of 34 patients with
primary site head and neck neoplasms were used. For each tumor, an experienced radiologist marked the lesion with a
bounding box and provided a reference standard by outlining the full 3D contour on both the pre- and post treatment
scans. A second radiologist independently marked each tumor again with another bounding box. The correlation between
the automatic and manual estimates for both the pre-to-post-treatment volume change and the percent volume change
was r=0.95. Based on the bounding boxes by the second radiologist, the correlation between the automatic and manual
estimate for the pre-to-post-treatment volume change was r=0.89 and for the percent volume change was r=0.91. The
correlation for the automatic estimates obtained from the bounding boxes by the two radiologists was as follows: (1) pretreatment
volume r=0.92, (2) post-treatment volume r=0.88, (3) pre-to-post-treatment change r=0.89 and (4) percent preto-
post-treatment change r=0.90. The difference between the automatic estimates based on the two sets of bounding
boxes did not achieve statistical significance for any of the estimates (p>0.29). The preliminary results indicate that the
automated segmentation system can reliably estimate tumor size change in response to treatment relative to radiologist's
hand segmentation as reference standard, and that the performance was robust against inter-observer variability in
marking the input bounding boxes.
|