Autocontouring Versus Manual Contouring

TO THE EDITOR: We read with interest the study of Wu et al. (1) regarding autocontouring methodologies for target delineation in PET/CT for non–small cell lung cancer (NSCLC). Seventeen NSCLC tumors were delineated with both automated and manual approaches, using either combined PET/CT or CT and PET independently. As expected, manual contouring of PET uptake correlated better with the maximum diameter of the primary tumor than did autocontouring using a fixed threshold at 50% of maximum tumor uptake. We believe that this result is largely associated with the various shortcomings of fixed-threshold approaches, a point that needs to be clearly emphasized. The authors have previously demonstrated that the best correlation between histopathology-derived maximum tumor diameters and image-derived ones was obtained using a 50% fixed threshold (2). This conclusion was reached by comparison with results obtained using other fixed-threshold values (from 20% to 55%), with a modest correlation of 0.77 and nonstatistically significant differences from the other fixed-threshold values tested. Most significantly, the use of a 50% fixed threshold led to differences larger than 1 cm in half the tumors considered. Such differences in maximum tumor diameter will most certainly lead to larger differences in the overall 3-dimensional volume. Considering similar comparisons based on 3-dimensional NSCLC tumor volumes determined by histopathology, other authors have demonstrated that an “optimal” threshold cannot be determined; considerable variability is seen (20%–42% [31% 6 11%] of the maximum), whereas CT-based volumes significantly overestimated the pathologic volume (3). It is therefore important to emphasize that a fixed threshold (irrespective of its absolute value) is not an adequate methodology to delineate elevated uptake signal in PET, because of its binary, deterministic nature and lack of robustness versus varying contrast and noise conditions (4,5). To account for these widely documented literature findings concerning tumor target delineation incorporating PET uptake information, fixed thresholding should be avoided, and at the very least, methodologies considering target-to-background ratios such as adaptive thresholding (5,6) should be favored. Eventually, the wider availability of automatic segmentation approaches (7–10), some of which can account for the presence of heterogeneous tumor uptake (7), may improve the accuracy and reproducibility of adaptive thresholding (11) for determination of functional tumor volume. Considering all these facts, we do agree with the authors that manual contouring should be preferred to autocontouring at a 50% threshold for functional tumor volume delineation. On the other hand, one should consider that manual delineation of PET uptake is not the ideal approach either, for multiple reasons. Most importantly, it represents a long process, particularly when it has to be performed in 3 dimensions, and it is inherently of low reproducibility (11). We therefore recommend that future studies investigating this issue include the use of advanced image segmentation approaches (4–10), which have demonstrated improved performance in comparison to a fixed threshold and may therefore lead to alternative or complementary conclusions regarding the role of manual contouring. Irrespective of the performance of a segmentation algorithm, operator intervention will always be necessary to appropriately identify the functional uptake of interest and avoid the inclusion of non–tumor-specific uptake.