ESTRO 37 Abstract book

S1190

ESTRO 37

Conclusion Atlas patients containing contour probability maps can provide a measure of accuracy for RADAR CTV contours. Investigations using a 50% probability map revealed DSC = 0.55 to be predictive of local regression, but not overall survival or rectal bleeding. Due to DSCs inability to differentiate between contours that lie within or outside the true volume, further analysis using additional contouring metrics may reveal more significant correlations. EP-2151 Personalized Deep Learning-Based Auto- Segmentation V. Kearney 1 , S. Haaf 2 , J. Chan 1 , S. Wu 1 , M. Bogdanov 1 , S. Yom 1 , T. Solberg 1 1 University of California UCSF, department of radiation oncology, San Francisco CA, USA 2 Nimble Therapy- LLC, Department of Artificial Intelligence, San Francisco, USA Purpose or Objective To demonstrate the feasibility of a personalized deep learning-based auto-segmentation system. Material and Methods Deep learning-based auto-segmentation was evaluated for the spinal cord, brainstem, left parotid, right parotid and mandible in 102 head and neck cancer patients. Each patient contains two ground-truth structure sets, one previously contoured in the clinic under normal working conditions (Human-1), the other carefully contoured by a single physician (Human-2). 3 different training variations of a deep learning-based automatic contouring system, Nimble Contour TM (Nimble Therapy, San Francisco, CA) were investigated using 13 validation patients and 89 training patients. 3 physicians and 1 dosimetrist participated in this study. Prior to training, the clinicians submitted scores for each contour on every patient in Human-1. Variations were trained on Human-2 (AI-1) and Human-1 (AI-3), and a third was trained on Human-1 incorporating the aggregated scores into the training phase (AI-2). For validation, each clinicians graded all human and AI data sets side by side, without knowledge of which data set they were grading (Figure 1). Scores ranging from 0 to 100 were assigned based on individual clinician preferences. Results Averaging over all the contours for all the patients, 1 clinician ranked AI-1 the highest and 3 clinicians ranked AI-1 the second highest (Table 1). 2 clinicians ranked AI-2 the 3 rd highest. On average AI-3 had the lowest performance but was only slightly lower than Human-2. Conclusion Clinician styles and preferences tend to vary considerably within a clinic. However, this study has demonstrated that by using clinician scoring of retrospective data, or carefully contoured prospective data, a deep learning auto-contouring model can be created that outperforms the original clinical data.

presented for the left parotid, right parotid, and spinal cord (B). The Human-2 (C), AI-1 (D), Human-1 (E), and AI- 2 (F) contours are presented for the mandible, (pink), left parotid (teal), and right parotid (orange). Note: Voxels above a prediction threshold of 0.5 are considered to be part of the structure under consideration.

Table 1. The clinician scores for the average of all 13 patients are presented for every OAR for each contouring style.

Electronic Poster: Physics track: Implementation of new technology, techniques, clinical protocols or trials (including QA & audit)

EP-2152 Validation Of A Novel Software For Correcting Distortion In Cranial Magnetic Resonance (Mr) Images J.F. Calvo Ortega 1 , J. Mateos 2 , A. Alberich 3 , S. Moragues 1 , J. Casals 1 1 HOSPITAL QUIRON BARCELONA, Radiotherapy, Barcelona, Spain 2 Imagen Ensayos Clínicos IEC- Hospital Quirón, Imagen Ensayos Clínicos IEC- Hospital Quirón, Barcelona, Spain 3 Biomedical Imaging Research Group GIBI230- La Fe Health Research Institute, Biomedical Imaging Research Group GIBI230- La Fe Health Research Institute, Valencia, Spain Purpose or Objective The accuracy of the Brainlab Elements Cranial Distortion software was investigated. Material and Methods Five MR datasets (used for vestibular schwannoma radiosurgery planning) were intentionally distorted. Two types of distortions were applied: 1) dist1: s=r(1+0.1r^2) and dist2: s= r(1+0.5r), where s and r are the distance from the center of distortion in the undistorted and distorted images, respectively. Each distorted MR dataset was corrected using the Cranial Distortion software, resulting a new corrected MR dataset (MRcorr). The accuracy of the correction was quantified by calculating the target registration error (TRE) for six anatomical landmarks identified in the co-registered MRcorr and planning CT (pCT) images. The chosen landmarks were points marked at of the two vestibules, at the two internal auditory canals and at the cochleas. The pCT dataset was selected as the reference to specify the 'true” position of each selected landmark. Results Figure 1 and 2 show the TRE values obtained for each type of forced distortion. In overall, the TRE values (0.6 mm ± 0.3 mm) were within the voxel size dimension of the pCT scan (1 x 1 x 1 mm 3 ). Legend of Figures 1 and 2: Red columns: average target registration error between the distorted MR set and the planning CT (TREdist). Blue columns: average target registration error between the corrected MR set and the planning CT (TREcorr). Mean ± standard deviation (SD) is displayed as a bar on each column of the plot. C: ipsilateral cochlea, V: ipsilateral vestibule, IAC: ipsilateral internal acoustic canal; Cc: contralateral cochlea, Vc: contralateral vestibule, IACc: contralateral internal acoustic canal.

Figure 1. The contour prediction map for AI-1 is presented for the left parotid, right parotid, and spinal cord (A). The Human-2 (H-2) and AI-1 contours are