ESTRO 38 Abstract book

S562 ESTRO 38

PO-1016 Segmentation of CT images with AI: compensating annotation uncertainties using contour augmentation U. Javaid 1 , D. Dasnoy 2 , J. Lee 1 1 UCLouvain, IREC/MIRO, Brussels, Belgium ; 2 UCLouvain, ICTeam, Louvain-la-neuve, Belgium Purpose or Objective Annotation of organs-at-risk ( OARs ) and target volumes ( TVs ) on CT scans is a key step in radiotherapy. Manual annotation is still in practice whereas to automate it, deep learning (DL) is currently being investigated. However, performance of DL networks ( DLNs ) depends highly on ground truth ( GT ) quality. For example, in radiotherapy, delineating TVs is of prime concern and proper care is taken while drawing them. OARs are not always a priority and sometimes roughly annotated. Rough annotation affects the predictive capabilities of DLNs in terms of organ localization. The network learning is thus sub- optimal, as it cannot improve over the given unique, possibly offset contour. Augmenting GT can address this issue. In this work, we propose contour augmentation to compensate the insufficient annotation quality by training a fully convolutional network ( dilated UNet ) with multiple GTs per image. Material and Methods To address annotation uncertainties, we augment the input contours per image and produce multiple contours. Given a single GT, we apply random deformation within 4mm to its x and y coordinates. A Gaussian smoothing is applied to smoothen the deformed contours to be qualitatively consistent with the original GT. Each deformed GT is considered as an additional GT. By incorporating multiple GT variations in the training phase, the model not only learns from a single ambiguous GT but from a sampled distribution of GTs, which can enhance its predictive capabilities, especially if some sampled GTs better correlate with features in the input image than the original GT. Our dataset consists of pelvic CT scans acquired from 67 patients (retrospective use). The CT images and their corresponding OAR annotations, in particular bladder and rectum were used for this study. We train our model on 1163 CT slices. Results After training, our model is able to take into account the different variabilities in GT and thus is more robust to missing/wrong annotations as compared to a model trained with a single GT. In the example test case (Figure 1), we show that our model is able to correct the missing annotation. We report Dice score, 95th percentile Hausdorff distance and average symmetric surface distance for bladder and rectum as 0.91±0.12 and 0.89±0.09, 2.08±0.93 and 1.38±1.09 mm , 0.24±0.11 and 0.14±0.08 mm respectively. Conclusion We address insufficient annotation quality by augmenting the manual annotations in GT. We call it contour augmentation. The proposed framework enables the DLNs to capture variability in terms of localizing the organ boundaries, allowing to correct the missing annotations at inference time. Our framework offers good generalization ability as it can be trained on any semantic segmentation task. We believe that contour augmentation can lead to accurate organ localization as it can cope with the possible bias introduced by annotation uncertainty.

Results With the modified applicator it is possible to achieve reasonably flat dose distribution in lateral and depth (< 6 mm) directions (Fig 2). After 6 mm depth there is a rapid dose fall-off with R50 = 1.0 cm and Rp = 1.6 cm. The penumbrae are sharp (±6 mm at depth of dose maximum). The energy filter plate reduces the dose output by a factor of 7.

Conclusion Our electron applicator has very favorable dose distribution characteristics for treating small superficial skin lesions, compared with either superficial x-ray devices or skin applicators of brachytherapy afterloaders. With x-rays < 100 kV tissues with higher density, like cartilage or bone, receive higher dose because of higher local absorption. Additionally, both x-rays and Ir-192 produce higher dose deeper than 5 mm in healthy tissue and higher dose inhomogeneity within the target depths ≤ 5 mm. Currently, we are initiating a clinical trial for testing the safety and efficacy of the new skin applicator.