Automatic detection of delineation outliers at an MR linac
MO-0879
Abstract
Automatic detection of delineation outliers at an MR linac
Authors: Carsten Brink1, Uffe Bernchou1, Irene Hazell1, Anders Bertelsen1, Ebbe L. Lorenzen1, Christian R. Hansen1, Rasmus L. Christiansen1, Nis Sarup1, Søren N. Agergaard1, Karina L. Gottlieb1, Henrik R. Jensen1, Rana Bahij2, Lars Dysager3, Christina J. Nyborg2, Olfred Hansen2, Tine Schytte2
1Laboratory of Radiation Physics, Department of Oncology, Odense University Hospital, Odense, Denmark; 2Department of Oncology, Odense University Hospital, Odense, Denmark; 3Department of Oncology, Odense University Hospital, Oden, Denmark
Show Affiliations
Hide Affiliations
Purpose or Objective
Using an MR linac, it is possible to adapt treatments based on MR images daily. For plan adaptation, structures are rigidly or deformably transferred from a reference to the MR of the day and manually corrected if needed. Since the patient is on the couch, these corrections are performed under time pressure, and therefore error-prone. The current study aims to develop and evaluate a machine learning system able to detect delineation outliers automatically.
Material and Methods
A cohort of 1629 treatment plans and 1453 unique MR scans related to MR linac treatments of 82 prostate cancer patients (60 Gy in 20 fractions) was used in this study. The pre-treatment plans of 40 randomly selected patients were used for model development, while their treatments were used for validation. Independent testing was performed in the test data from all treatments of the remaining 42 patients.
The outlier detection system used the training data to build a library of typical variations based on delineation information such as relative centre-of-mass position, volume, length, area, and spatial direction. The algorithm predicted the likelihood that new delineations were similar to those in the library, and delineations with low likelihood were categorised as outliers. An oncologist classified outliers as true or false positive based on the question, “would you have corrected this if you had been aware of it at treatment time?”.
If needed, the transferred delineations were either manually edited or re-delineated on a subset of images followed by interpolation. Delineations were typically only corrected within a delineation volume (e.g. 2 cm around PTV) to balance the need for a quick delineation procedure versus the needs of the adaptive treatment planning process. Thus, outliers that related to intentionally uncorrected delineations were classified as false-positive.
Results
Outliers were detected in 133 image sets (67 in validation data/66 in test data). Some outliers were related to the same issue repeated in multiple treatment fractions. Excluding these “double” counts, 28 unique outliers (13 in validation data/15 in test data) were identified. Thus unique outliers were only detected in 2% (28/1453) of the image sets. However, due to the multiple fraction treatment course, true positive outliers were located for approximately 20% of the patients. Examples of identified outliers were: interchange of left and right femoral head, missing contour interpolations (fig. 1), overlapping delineations (fig. 2), and deformed contours that lacked correction inside delineation volume. Of the 28 outliers, 53% and 73% were classified as true-positive observations for the validation and test data, respectively.
Conclusion
The machine learning system was able to find relevant
deviations based only on local delineation practice. More than half of the
outliers were classified as true outliers. Based on these results, clinical
implementation at the MR linac has been initiated.