Automated clinical target volume delineation using deep 3D neural networks in radiation therapy of Non-small Cell Lung Cancer

Abstract

Background and purpose Clinical targeted volume (CTV) delineation accounting for the patient-specific microscopic tumor spread can be a difficult step in defining the treatment volume. We developed an intelligent and automated CTV delineation system for locally advanced non-small cell lung carcinoma (NSCLC) to cover the microscopic tumor spread while avoiding organs-at-risk (OAR). Materials and methods A 3D UNet with a customized loss function was used, which takes both the patients’ respiration-correlated (“4D”) CT scan and the physician contoured internal gross target volume (iGTV) as inputs, and outputs the CTV delineation. Among the 84 identified patients, 60 were randomly selected to train the network, and the remaining as testing. The model performance was evaluated and compared with cropped expansions using the shape similarities to the physicians’ contours (the ground-truth) and the avoidance of critical OARs. Results On the testing datasets, all model-predicted CTV contours followed closely to the ground truth, and were acceptable by physicians. The average dice score was 0.86. Our model-generated contours demonstrated better agreement with the ground-truth than the cropped 5 mm/8 mm expansion method (median of median surface distance of 1.0 mm vs 1.9 mm/2.0 mm), with a small overlap volume with OARs (0.4 cm 3 for the esophagus and 1.2 cm 3 for the heart). Conclusions The CTVs generated by our CTV delineation system agree with the physician's contours. This approach demonstrates the capability of intelligent volumetric expansions with the potential to be used in clinical practice.

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

“…Recent studies have elicited dose–response relationships for several component structures [15] , [16] , [17] , [18] , [19] , [20] and trials are underway to test the effect of sparing these regions [30] . As the application of artificial intelligence is explored in RT, volume delineation serves as a logical starting point [31] , [32] , [33] . By accurately performing complex and time-consuming delineation tasks, clinician availability for alternate activities could be increased and treatment planning delays reduced [34] .…”

A pair of deep learning auto‐contouring models for prostate cancer patients injected with a radio‐transparent versus radiopaque hydrogel spacer

“…17 Kawula et al trained a model using an in-house algorithm, and demonstrated DSC of 0.87, 0.97, and 0.89 for prostate, bladder, and rectum, respectively. 18 In addition to prostate cancer, CT-based DL auto-segmentation models have also been reported for other disease sites, such as brain, 19,20 head and neck, 21,22 lung, [23][24][25][26] breast, 27 esophagus, 28 liver, [29][30][31] pancreas, 32 bladder, 33 and gynecological cancers. 34,35 While the prior studies have demonstrated the feasibility of constructing DL models to automate the tedious contouring task in prostate treatment planning, the applicability of each model is limited by clinical protocol, particularly the immobilization method which can affect the shape and position of key organs.…”

“…trained a model using an in‐house algorithm, and demonstrated DSC of 0.87, 0.97, and 0.89 for prostate, bladder, and rectum, respectively 18 . In addition to prostate cancer, CT‐based DL auto‐segmentation models have also been reported for other disease sites, such as brain, 19,20 head and neck, 21,22 lung, 23–26 breast, 27 esophagus, 28 liver, 29–31 pancreas, 32 bladder, 33 and gynecological cancers 34,35 …”

Towards real-time radiotherapy planning: The role of autonomous treatment strategies

“…The one-day workflow proposed by Palacios et al [2] used automation only to a minor extent and it is therefore highly dependent on the availability of staff throughout the day and not easily scalable to increasing patient numbers. Automated tools for various steps in the radiotherapy planning workflow such as automatic contouring [4] , [5] , [6] , [7] , [8] and radiotherapy planning [9] , [10] , [11] , [12] , [13] recently gained attention. For instance, Johnston et al [7] showed the usability of a convolutional neural network for segmentation of thoracic organs at risk.…”

“…For instance, Johnston et al [7] showed the usability of a convolutional neural network for segmentation of thoracic organs at risk. Although auto-contouring of targets is more challenging, Xie et al [8] recently introduced a 3D neural network for lung lesion contouring. Also, for treatment plan optimization different approaches were proposed [9] , [10] , [11] , [12] , [13] .…”