ESTRO 38 Abstract book

S1135 ESTRO 38

assignment for two tissue types is a feasible approach for patients with prostate cancer. E P-2062 Feasibility of automatic detection of breast limits for auto-planning J. Oliveira 1 , M.C. Aznar 2 , Y. Kirova 3 , A. Henry 3 , P. Aljabar 1 , M. Van Herk 2 , P. Poortmans 3 , M. Gooding 1 1 Mirada Medical Ltd, Science department, Oxford, United Kingdom ; 2 University of Manchester, Division of Cancer Sciences, Manchester, United Kingdom ; 3 Institut Curie, Radiation Oncology, Paris, France Purpose or Objective The planning of tangential breast radiotherapy requires the detection of the limits of the breast on CT imaging. The purpose of this abstract is to assess the feasibility of automated methods for detecting those limits without resorting to any manual annotations or markers (e.g., Three automated methods were evaluated in this work. The first used a convolutional neural network (CNN) to predict the slice coordinate in a reference volume given a single slice of a query volume (e.g., finding slice location in the superior-inferior axis). Once these normalized coordinates are computed for the query volume, the breast limits can be predicted by matching the normalized coordinates with manual annotations of the limits in the reference volume. The second method used a CNN to predict the normalized distance to the selected limits given an input slice. This approach is more focused on the limits, nonetheless it has the downside of requiring manual annotations for all the volumes in the training set. The ground truth for training the CNNs of these two approaches was derived from the deformable registration of the reference volume to the query volumes. The third method predicts the limits of breast clinical target volumes (CTVs) through atlas-based segmentation. This is a simpler approach that requires less parameter tuning, but it requires manually contouring the CTVs for the whole training set. The methods were evaluated in a database of 56 CT scans from breast cancer patients, including manual delineations of breast CTVs contoured according to the ESTRO guidelines. The delineations were used to define the cranial and caudal limits of each CTV. One CT scan was defined as the reference, while the remaining database was divided into 5 folds, where 4 were used for training and 1 for testing. The methods were evaluated on the 55 CT scans by rotating the training and test sets across folds. Results The results for detecting the cranial and caudal limits of breast CTVs are presented in Table 1. Method 2 outperformed method 1 (Figure 1), which may be justified with the former method being more focused on detecting the limits, allowing the CNN to learn a more direct mapping between image features and breast limits. Methods 2 and 3 achieved comparable results, with method 2 performing better for the caudal limit but worse for the cranial limit. wires) for the patient. Material and Methods

Material and Methods CT and MR images were acquired in the treatment position for five prostate cancer patients. The planning target and OARs were delineated according to the standard practice of the clinic. Conventional VMAT plans were generated for each patient based on the CT image. In order to create a tissue density map to calculate dose on the MR image, a bone ROI was created by a multi-atlas-based segmentation algorithm. To minimise the effect of anatomical differences between the CT and MR image, the external contour defined on the CT image was used for the MR image. The CT based plan was recalculated on the MR image by assigning bulk densities to the external contour and bone. The calculated dose for the CT based plan was compared to the recalculated dose based on the MR image. In order to find the optimal bulk density assignments for soft tissue and bone, a range of soft-tissue and bone density pairs were evaluated for all patients. The soft- tissue density varied from 0.95-1.03 g/cm 3 in 0.01 increments and the bone density varied from 1.15 - 1.65 g/cm 3 in 0.05 increments. Dose differences between the CT and the bulk density based dose distributions were calculated for D1, D2, D50, D95, D98, D99 and average dose for the PTV, Rectum and Bladder. The density pair with the lowest dose difference between the CT and MR based calculations among all dose statistics and for all ROIs and patients was considered the best and used for the rest of the study. Further, the robustness of this bulk- density approach was investigated by looking at the average dose per ROI for all five patients. Results With a mean absolute dose difference of 0.3% of the prescribed dose over all patients, ROIs and dose statistics, the best results were found with a density of 0.98 g/cm 3 for soft tissue and 1.25 g/cm 3 for bone. In Table 1 the average percentage dose difference of the prescribed dose is shown for each ROI and patient with these densities. The mean average dose was close to 0.0% and the standard deviation was 0.5% or less for all ROIs. Figure 1 shows the CT based dose distribution, MR based dose distribution and DVHs for PTV, Rectum and Bladder for Patient 1.

Conclusion Using bulk-density assignment on MR images with dosimetrically optimized densities for bone and soft tissue results in small dose differences compared to dose calculated on the CT. This demonstrates that an integrated MR-only pathway utilizing a bulk density