Room: Exhibit Hall | Forum 6
Purpose: To evaluate the correlation between Dice similarity coefficient (DSC) and treatment plan metrics in order to judge its suitability for auto-contouring of anatomy for treatment planning purposes.
Methods: DSC has been widely used for evaluation of the similarity between two contours. It is defined as two times the overlap volume between two contours divided by the total volume of the two individual contours. 17 subjects were contoured by physicians and autocontouring software based on the machine learning method of decision forests. The software was initially trained on randomly selected 94 CT images and is able to autocontour prostate GTV, seminal vesicles (SV), bladder, rectum and femurs. After generating the autocontours, we then reproduced additional contours used during optimization, such as PTV, based on physiciansâ€™ instructions for each patient. New plans used the same beam angles and optimization objectives as were used for the clinical plans. Dosimetry for both the original plan and the autocontouring plan was calculated on the physicianâ€™s original â€˜gold standardâ€™ contours. Metrics based on physiciansâ€™ clinical goals, including PTV-D95%, rectal wall V75Gy, and bladder wall V80Gy, were used to evaluate the dose distributions of the two plans and study their correlations with their DSC.
Results: The PTV-D95% metric showed a good correlation with the associated DSC. For lower values of the DSC the correlation breaks down. For OAR metrics which tend to be volume-metrics, instead of dose-metrics, the relationship between the DCS is more complex.
Conclusion: Due to the fact that the treatment plans essentially matched the isodose lines to the prostate PTV contours, the DSC accurately predicted the plan quality for the target. The DSC has some limited use in evaluating contouring algorithms for therapy purposes.