Deep learning model for automatic contouring of cardiovascular substructures on radiotherapy planning CT images: Dosimetric validation and reader study based clinical acceptability testing

Abstract

Background and purpose: Large radiotherapy (RT) planning imaging datasets with consistently contoured cardiovascular structures are essential for robust cardiac radiotoxicity research in thoracic cancers. This study aims to develop and validate a highly accurate automatic contouring model for the heart, cardiac chambers, and great vessels for RT planning computed tomography (CT) images that can be used for dose-volume parameter estimation. Materials and methods: A neural network model was trained using a dataset of 127 expertly contoured planning CT images from RT treatment of locally advanced non-small-cell lung cancer (NSCLC) patients. Evaluation of geometric accuracy and quality of dosimetric parameter estimation was performed on 50 independent scans with contrast and without contrast enhancement. The model was further evaluated regarding the clinical acceptability of the contours in 99 scans randomly sampled from the RTOG-0617 dataset by three experienced radiation oncologists. Results: Median surface dice at 3 mm tolerance for all dedicated thoracic structures was 90% in the test set. Median absolute difference between mean dose computed with model contours and expert contours was 0.45 Gy averaged over all structures. The mean clinical acceptability rate by majority vote in the RTOG-0617 scans was 91%. Conclusion: This model can be used to contour the heart, cardiac chambers, and great vessels in large datasets of RT planning thoracic CT images accurately, quickly, and consistently. Additionally, the model can be used as a time-saving tool for contouring in clinic practice.

Validation of an established deep learning auto-segmentation tool for cardiac substructures in 4D radiotherapy planning scans

“…Two alternate auto-segmentation algorithms that use a neural network architecture have been published, subsequent to the development of the tool presented, trained using 3D-CT planning cases. The model published by Garrett Fernandes et al had DSC of 0.74–0.95 across the same set of substructures and the largest median absolute difference in mean doses in the range 0.1 Gy–1.0 Gy [35] . The model published by Van Velzen et al had a DSC of 0.76–0.88 and R 2 values for dosimetric parameters were 0.77–1.00 [53] .…”

“…In this study, a deep learning-based cardiac substructure auto-segmentation tool developed for use in 3D-CT scans was retrospectively evaluated in 20 patients that underwent 4D-CT planning. This particular tool was selected from the literature [35] , [36] , [37] , [38] , [39] , [40] , [41] , [42] , [43] , [44] , [45] , [46] due to its superiority in terms of structures included, performance metrics, applicability to lung cancer and availability. The cardiac substructure auto-segmentation available tools at the time of writing are compared in Supplementary Table 11 .…”

Open-source, fully-automated hybrid cardiac substructure segmentation: development and optimisation

“…Compared to many other studies presenting models for automatic segmentation of cardiac substructures, this work uses far less training data to achieve similar results, requiring only 10 images with manually contoured cardiac substructures compared to other approaches which used 41 [ 19 ], 127 [ 50 ], and 217 [ 14 ] cases for training. An overview of published tools for cardiac substructure segmentation can be found in a recent publication by Walls et al (Supplementary Table 11) [ 18 ].…”

“…This method has also been validated on an independent dataset [ 18 ], with results suggesting reduced accuracy when applied to new data as well as systematic variations. A DL model developed by Garrett Fernandes et al [ 50 ] to delineate these same cardiac substructures from a training dataset 127 CT scans was validated on an independent dataset and also achieved higher DSC values, however a substantial reduction in performance was observed on CT imaging acquired without contrast enhancement. A cascading deep learning model was recently proposed by van den Oever [ 19 ] which automatically segments the heart and chambers accurately, although this method was only tested on 6 patient CT scans.…”

The impact of inter-observer variation in delineation on robustness of radiomics features in non-small cell lung cancer

“…All patients had a spiral CT scan taken of 3 mm slice thickness (voxel dimensions 1 mm × 1 mm × 3 mm). The majority of CT scans are non-contrast scans, as per a recently published study that provided a quantitative analysis which classifies scans from this dataset as contrast or non-contrast 29 . At the time of initiation of the study (October 2018 to January 2019), radiotherapy structure sets were not available for 92 patients on the TCIA website and these were manually contoured at Peter MacCallum Cancer Centre (PMCC) by a radiation oncologist (GK) with additional review by a lung radiologist (BW) as required.…”