A convolutional neural network algorithm for automatic segmentation of head and neck organs at risk using deep lifelong learning.

PURPOSE: This study suggests a lifelong learning-based convolutional neural network (LL-CNN) algorithm as a superior alternative to single-task learning approaches for automatic segmentation of head and neck (OARs) organs at risk. METHODS AND MATERIALS: Lifelong learning-based convolutional neural network was trained on twelve head and neck OARs simultaneously using a multitask learning framework. Once the weights of the shared network were established, the final multitask convolutional layer was replaced by a single-task convolutional layer. The single-task transfer learning network was trained on each OAR separately with early stoppage. The accuracy of LL-CNN was assessed based on Dice score and root-mean-square error (RMSE) compared to manually delineated contours set as the gold standard. LL-CNN was compared with 2D-UNet, 3D-UNet, a single-task CNN (ST-CNN), and a pure multitask CNN (MT-CNN). Training, validation, and testing followed Kaggle competition rules, where 160 patients were used for training, 20 were used for internal validation, and 20 in a separate test set were used to report final prediction accuracies. RESULTS: On average contours generated with LL-CNN had higher Dice coefficients and lower RMSE than 2D-UNet, 3D-Unet, ST- CNN, and MT-CNN. LL-CNN required ~72 hrs to train using a distributed learning framework on 2 Nvidia 1080Ti graphics processing units. LL-CNN required 20 s to predict all 12 OARs, which was approximately as fast as the fastest alternative methods with the exception of MT-CNN. CONCLUSIONS: This study demonstrated that for head and neck organs at risk, LL-CNN achieves a prediction accuracy superior to all alternative algorithms.

Use of a realistic breathing lung phantom to evaluate dose delivery errors.

PURPOSE: To compare the effect of respiration-induced motion on delivered dose (the interplay effect) for different treatment techniques under realistic clinical conditions. METHODS: A flexible resin tumor model was created using rapid prototyping techniques based on a computed tomography (CT) image of an actual tumor. Twenty micro-MOSFETs were inserted into the tumor model and the tumor model was inserted into an anthropomorphic breathing phantom. Phantom motion was programed using the motion trajectory of an actual patient. A four-dimensional CT image was obtained and several treatment plans were created using different treatment techniques and planning systems: Conformal (Eclipse), step-and-shoot intensity-modulated radiation therapy (IMRT) (Pinnacle), step-and-shoot IMRT (XiO), dynamic IMRT (Eclipse), complex dynamic IMRT (Eclipse), hybrid IMRT [60% conformal, 40% dynamic IMRT (Eclipse)], volume-modulated are therapy (VMAT) [single-arc (Eclipse)], VMAT [double-arc (Eclipse)], and complex VMAT (Eclipse). The complex plans were created by artificially pushing the optimizer to give complex multileaf collimator sequences. Each IMRT field was irradiated five times and each VMAT field was irradiated ten times, with each irradiation starting at a random point in the respiratory cycle. The effect of fractionation was calculated by randomly summing the measured doses. The maximum deviation for each measurement point per fraction and the probability that 95% of the model tumor had dose deviations less than 2% and 5% were calculated as a function of the number of fractions. Tumor control probabilities for each treatment plan were calculated and compared. RESULTS: After five fractions, measured dose deviations were less than 2% for more than 95% of measurement points within the tumor model for all plans, except the complex dynamic IMRT, step-and-shoot IMRT (XiO), complex VMAT, and single-arc VMAT plans. Reducing the dose rate of the complex IMRT plans from 600 to 200 MU/min reduced the dose deviations to less than 2%. Dose deviations were less than 5% after five fractions for all plans, except the complex single-arc VMAT plan. CONCLUSIONS: Rapid prototyping techniques can be used to create realistic tumor models. For most treatment techniques, the dose deviations averaged out after several fractions. Treatments with unusually complicated multileaf collimator sequences had larger dose deviations. For IMRT treatments, dose deviations can be reduced by reducing the dose rate. For VMAT treatments, using two arcs instead of one is effective for reducing dose deviations.

Use of reduced dose rate when treating moving tumors using dynamic IMRT.

The purpose was to evaluate the effect of dose rate on discrepancies between expected and delivered dose caused by the interplay effect. Fifteen separate dynamic IMRT plans and five hybrid IMRT plans were created for five patients (three IMRT plans and one hybrid IMRT plan per patient). The impact of motion on the delivered dose was evaluated experimentally for each treatment field for different dose rates (200 and 400 MU/min), and for a range of target amplitudes and periods. The maximum dose discrepancy for dynamic IMRT fields was 18.5% and 10.3% for dose rates of 400 and 200 MU/min, respectively. The maximum dose discrepancy was larger than this for hybrid plans, but the results were similar when weighted by the contribution of the IMRT fields. The percentage of fields for which 98% of the target never experienced a 5% or 10% dose discrepancy increased when the dose rate was reduced from 400 MU/min to 200 MU/min. For amplitudes up to 2 cm, reducing the dose rate to 200 MU/min is effective in keeping daily dose discrepancies for each field within 10%.

Hybrid IMRT for treatment of cancers of the lung and esophagus.

PURPOSE: To report on a hybrid intensity-modulated radiation therapy (IMRT; static plus IMRT beams treated concurrently) technique for lung and esophageal patients to reduce the volume of lung treated to low doses while delivering a conformal dose distribution. METHODS: Treatment plans were analyzed for 18 patients (12 lung and 6 esophageal). Patients were treated with a hybrid technique that concurrently combines static (approximately two-thirds dose) and IMRT (approximately one-third dose) beams. These plans were compared with conventional three-dimensional (3D; non-IMRT) plans and all IMRT plans using custom four- and five-field arrangements and nine equally spaced coplanar beams. Plans were optimized to reduce V13 and V5 values. Dose-volume histograms were calculated for the planning target volume, heart, and the ipsilateral, contralateral, and total lung. Lung volumes V5, V13, V20, V30; mean lung dose (MLD); and the generalized equivalent uniform dose (gEUD) were calculated for each plan. RESULTS: Hybrid plans treated significantly smaller total and contralateral lung volumes with low doses than nine-field IMRT plans. Largest reductions were for contralateral lung V5, V13, and V20 values for lung (-11%, -15%, -7%) and esophageal (-16%, -20%, -7%) patients. Smaller reductions were found also for 3D and four- and five-field IMRT plans. MLD and gEUDs were similar for all plan types. The 3D plans treated much larger extra planning target volumes to prescribed dose levels. CONCLUSIONS: Hybrid IMRT demonstrated advantages for reduction of low-dose lung volumes in the thorax for reducing low dose to lung while also reducing the potential magnitude of dose deviations due to intrafraction motion and small field calculation accuracy.