U-net architecture with embedded Inception-ResNet-v2 image encoding modules for automatic segmentation of organs-at-risk in head and neck cancer radiation therapy based on computed tomography scans.

Purpose.The purpose of this study was to utilize a deep learning model with an advanced inception module to automatically contour critical organs on the computed tomography (CT) scans of head and neck cancer patients who underwent radiation therapy treatment and interpret the clinical suitability of the model results through activation mapping.Materials and methods.This study included 25 critical organs that were delineated by expert radiation oncologists. Contoured medical images of 964 patients were sourced from a publicly available TCIA database. The proportion of training, validation, and testing samples for deep learning model development was 65%, 25%, and 10% respectively. The CT scans and segmentation masks were augmented with shift, scale, and rotate transformations. Additionally, medical images were pre-processed using contrast limited adaptive histogram equalization to enhance soft tissue contrast while contours were subjected to morphological operations to ensure their structural integrity. The segmentation model was based on the U-Net architecture with embedded Inception-ResNet-v2 blocks and was trained over 100 epochs with a batch size of 32 and an adaptive learning rate optimizer. The loss function combined the Jaccard Index and binary cross entropy. The model performance was evaluated with Dice Score, Jaccard Index, and Hausdorff Distances. The interpretability of the model was analyzed with guided gradient-weighted class activation mapping.Results.The Dice Score, Jaccard Index, and mean Hausdorff Distance averaged over all structures and patients were 0.82 ± 0.10, 0.71 ± 0.10, and 1.51 ± 1.17 mm respectively on the testing data sets. The Dice Scores for 86.4% of compared structures was within range or better than published interobserver variability derived from multi-institutional studies. The average model training time was 8 h per anatomical structure. The full segmentation of head and neck anatomy by the trained network required only 6.8 s per patient.Conclusions.High accuracy obtained on a large, multi-institutional data set, short segmentation time and clinically-realistic prediction reasoning make the model proposed in this work a feasible solution for head and neck CT scan segmentation in a clinical environment.

Adaptive radiation therapy strategies in the treatment of prostate cancer patients using hypofractionated VMAT.

PURPOSE: To perform a comprehensive evaluation of eight adaptive radiation therapy strategies in the treatment of prostate cancer patients who underwent hypofractionated volumetric modulated arc therapy (VMAT) treatment. MATERIAL AND METHODS: The retrospective study included 20 prostate cancer patients treated with 40 Gy total dose over five fractions (8 Gy/fraction) using VMAT. Daily cone beam computed tomography images were acquired before the delivery of every fraction and then, with the application of deformable image registration used for the estimation of daily dose, contouring and plan re-optimization. Dosimetric benefits of the various ART strategies were quantified by the comparison of dose and dose-volume metrics derived from treatment planning objectives for original treatment plan and adapted plans with the consideration of target volumes (PTV and CTV) as well as critical structures (bladder, rectum, left, and right femoral heads). RESULTS: Percentage difference (ΔD) between planning objectives and delivered dose in the D99%  > 4000cGy (CTV) metric was -3.9% for the non-ART plan and 2.1% to 4.1% for ART plans. For D99%  > 3800cGy and Dmax  < 4280cGy (PTV), ΔD was -11.2% and -6.5% for the non-ART plan as well as -3.9% to -1.6% and -0.2% to 1.8% for ART plans, respectively. For D15%  < 3200 cGy and D20%  < 2800 cGy (bladder), ΔD was -62.4% and -68.8% for the non-ART plan as well as -60.0% to -57.4% and -67.0% to -64.0% for ART plans. For D15%  < 3200 cGy and D20%  < 2800 cGy (rectum), ΔD was -11.4% and -8.15% for non-ART plan as well as -14.9% to -9.0% and -11.8% to -5.1% for ART plans. CONCLUSIONS: Daily on-line adaptation approaches were the most advantageous, although strategies adapting every other fraction were also impactful while reducing relative workload as well. Offline treatment adaptations were shown to be less beneficial due to increased dose delivered to bladder and rectum compared toother ART strategies.

Machine learning for dose-volume histogram based clinical decision-making support system in radiation therapy plans for brain tumors.

PURPOSE: To create and investigate a novel, clinical decision-support system using machine learning (ML). METHODS AND MATERIALS: The ML model was developed based on 79 radiotherapy plans of brain tumor patients that were prescribed a total dose of 60 Gy delivered with volumetric-modulated arc therapy (VMAT). Structures considered for analysis included planning target volume (PTV), brainstem, cochleae, and optic chiasm. The model aimed to classify the target variable that included class-0 corresponding to plans for which the PTV treatment planning objective was met and class-1 that was associated with plans for which the PTV objective was not met due to the priority trade-off to meet one or more organs-at-risk constraints. Several models were evaluated using double-nested cross-validation and an area-under-the-curve (AUC) metric, with the highest performing one selected for further investigation. The model predictions were explained with Shapely additive explanation (SHAP) interaction values. RESULTS: The highest-performing model was Logistic Regression achieving an accuracy of 93.8 ± 4.1% and AUC of 0.98 ± 0.02 on the testing data. The SHAP analysis indicated that the ΔD99% metric for PTV had the greatest influence on the model predictions. The least important feature was ΔDMAX for the left and right cochleae. CONCLUSIONS: The trained model achieved satisfactory accuracy and can be used by medical physicists in a data-driven quality assurance program as well as by radiation oncologists to support their decision-making process in terms of treatment plan approval and potential plan modifications. Model explanation analysis showed that the model relies on clinically valid logic when making predictions.