Hippocampus segmentation in CT using deep learning: impact of MR versus CT-based training contours.

Purpose: Hippocampus contouring for radiotherapy planning is performed on MR image data due to poor anatomical visibility on computed tomography (CT) data. Deep learning methods for direct CT hippocampus auto-segmentation exist, but use MR-based training contours. We investigate if these can be replaced by CT-based contours without loss in segmentation performance. This would remove the MR not only from inference but also from training. Approach: The hippocampus was contoured by medical experts on MR and CT data of 45 patients. Convolutional neural networks (CNNs) for hippocampus segmentation on CT were trained on CT-based or propagated MR-based contours. In both cases, their predictions were evaluated against the MR-based contours considered as the ground truth. Performance was measured using several metrics, including Dice score, surface distances, and contour Dice score. Bayesian dropout was used to estimate model uncertainty. Results: CNNs trained on propagated MR contours (median Dice 0.67) significantly outperform those trained on CT contours (0.59) and also experts contouring manually on CT (0.59). Differences between the latter two are not significant. Training on MR contours results in lower model uncertainty than training on CT contours. All contouring methods (manual or CNN) on CT perform significantly worse than a CNN segmenting the hippocampus directly on MR (median Dice 0.76). Additional data augmentation by rigid transformations improves the quantitative results but the difference remains significant. Conclusions: CT-based training contours for CT hippocampus segmentation cannot replace propagated MR-based contours without significant loss in performance. However, if MR-based contours are used, the resulting segmentations outperform experts in contouring the hippocampus on CT.

Clinical evaluation of a full-image deep segmentation algorithm for the male pelvis on cone-beam CT and CT.

AIM: The segmentation of organs from a CT scan is a time-consuming task, which is one hindrance for adaptive radiation therapy. Through deep learning, it is possible to automatically delineate organs. Metrics like dice score do not necessarily represent the impact for clinical practice. Therefore, a clinical evaluation of the deep neural network is needed to verify the segmentation quality. METHODS: In this work, a novel deep neural network is trained on 300 CT and 300 artificially generated pseudo CBCTs to segment bladder, prostate, rectum and seminal vesicles from CT and cone beam CT scans. The model is evaluated on 45 CBCT and 5 CT scans through a clinical review performed by three different clinics located in Europe, North America and Australia. RESULTS: The deep learning model is scored either equally good (prostate and seminal vesicles) or better (bladder and rectum) than the structures from routine clinical practice. No or minor corrections are required for 97.5% of the segmentations of the bladder, 91.5% of the prostate, 94% of the rectum and seminal vesicles. Overall, for 82.5% of the patients none of the organs need major corrections or a redraw. CONCLUSION: This study shows that modern deep neural networks are capable of producing clinically applicable organ segmentation for the male pelvis. The model is able to produce acceptable structures as frequently as current clinical routine. Therefore, deep neural networks can simplify the clinical workflow by offering initial segmentations. The study further shows that to retain the clinicians' personal preferences a structure review and correction is necessary for structures both created by other clinicians and deep neural networks.

Evaluation of deep learning methods for parotid gland segmentation from CT images.

The segmentation of organs at risk is a crucial and time-consuming step in radiotherapy planning. Good automatic methods can significantly reduce the time clinicians have to spend on this task. Due to its variability in shape and low contrast to surrounding structures, segmenting the parotid gland is challenging. Motivated by the recent success of deep learning, we study the use of two-dimensional (2-D), 2-D ensemble, and three-dimensional (3-D) U-Nets for segmentation. The mean Dice similarity to ground truth is ∼ 0.83 for all three models. A patch-based approach for class balancing seems promising for false-positive reduction. The 2-D ensemble and 3-D U-Net are applied to the test data of the 2015 MICCAI challenge on head and neck autosegmentation. Both deep learning methods generalize well onto independent data (Dice 0.865 and 0.88) and are superior to a selection of model- and atlas-based methods with respect to the Dice coefficient. Since appropriate reference annotations are essential for training but often difficult and expensive to obtain, it is important to know how many samples are needed for training. We evaluate the performance after training with different-sized training sets and observe no significant increase in the Dice coefficient for more than 250 training cases.

American Association of Physicists in Medicine Task Group 263: Standardizing Nomenclatures in Radiation Oncology.

A substantial barrier to the single- and multi-institutional aggregation of data to supporting clinical trials, practice quality improvement efforts, and development of big data analytics resource systems is the lack of standardized nomenclatures for expressing dosimetric data. To address this issue, the American Association of Physicists in Medicine (AAPM) Task Group 263 was charged with providing nomenclature guidelines and values in radiation oncology for use in clinical trials, data-pooling initiatives, population-based studies, and routine clinical care by standardizing: (1) structure names across image processing and treatment planning system platforms; (2) nomenclature for dosimetric data (eg, dose-volume histogram [DVH]-based metrics); (3) templates for clinical trial groups and users of an initial subset of software platforms to facilitate adoption of the standards; (4) formalism for nomenclature schema, which can accommodate the addition of other structures defined in the future. A multisociety, multidisciplinary, multinational group of 57 members representing stake holders ranging from large academic centers to community clinics and vendors was assembled, including physicists, physicians, dosimetrists, and vendors. The stakeholder groups represented in the membership included the AAPM, American Society for Radiation Oncology (ASTRO), NRG Oncology, European Society for Radiation Oncology (ESTRO), Radiation Therapy Oncology Group (RTOG), Children's Oncology Group (COG), Integrating Healthcare Enterprise in Radiation Oncology (IHE-RO), and Digital Imaging and Communications in Medicine working group (DICOM WG); A nomenclature system for target and organ at risk volumes and DVH nomenclature was developed and piloted to demonstrate viability across a range of clinics and within the framework of clinical trials. The final report was approved by AAPM in October 2017. The approval process included review by 8 AAPM committees, with additional review by ASTRO, European Society for Radiation Oncology (ESTRO), and American Association of Medical Dosimetrists (AAMD). This Executive Summary of the report highlights the key recommendations for clinical practice, research, and trials.