A novel CT-based radiomics model for predicting response and prognosis of chemoradiotherapy in esophageal squamous cell carcinoma.

No clinically relevant biomarker has been identified for predicting the response of esophageal squamous cell carcinoma (ESCC) to chemoradiotherapy (CRT). Herein, we established a CT-based radiomics model with artificial intelligence (AI) to predict the response and prognosis of CRT in ESCC. A total of 44 ESCC patients (stage I-IV) were enrolled in this study; training (n = 27) and validation (n = 17) cohorts. First, we extracted a total of 476 radiomics features from three-dimensional CT images of cancer lesions in training cohort, selected 110 features associated with the CRT response by ROC analysis (AUC ≥ 0.7) and identified 12 independent features, excluding correlated features by Pearson's correlation analysis (r ≥ 0.7). Based on the 12 features, we constructed 5 prediction models of different machine learning algorithms (Random Forest (RF), Ridge Regression, Naive Bayes, Support Vector Machine, and Artificial Neural Network models). Among those, the RF model showed the highest AUC in the training cohort (0.99 [95%CI 0.86-1.00]) as well as in the validation cohort (0.92 [95%CI 0.71-0.99]) to predict the CRT response. Additionally, Kaplan-Meyer analysis of the validation cohort and all the patient data showed significantly longer progression-free and overall survival in the high-prediction score group compared with the low-prediction score group in the RF model. Univariate and multivariate analyses revealed that the radiomics prediction score and lymph node metastasis were independent prognostic biomarkers for CRT of ESCC. In conclusion, we have developed a CT-based radiomics model using AI, which may have the potential to predict the CRT response as well as the prognosis for ESCC patients with non-invasiveness and cost-effectiveness.

Virtual cone-beam computed tomography simulator with human phantom library and its application to the elemental material decomposition.

PURPOSE: The purpose of this study is to develop a virtual CBCT simulator with a head and neck (HN) human phantom library and to demonstrate the feasibility of elemental material decomposition (EMD) for quantitative CBCT imaging using this virtual simulator. METHODS: The library of 36 HN human phantoms were developed by extending the ICRP 110 adult phantoms based on human age, height, and weight statistics. To create the CBCT database for the library, a virtual CBCT simulator that simulated the direct and scattered X-ray on a flat panel detector using ray-tracing and deep-learning (DL) models was used. Gaussian distributed noise was also included on the flat panel detector, which was evaluated using a real CBCT system. The usefulness of the virtual CBCT system was demonstrated through the application of the developed DL-based EMD model for case involving virtual phantom and real patient. RESULTS: The virtual simulator could generate various virtual CBCT images based on the human phantom library, and the prediction of the EMD could be successfully performed by preparing the CBCT database from the proposed virtual system, even for a real patient. The CBCT image degradation owing to the scattered X-ray and the statistical noise affected the prediction accuracy, although these effects were minimal. Furthermore, the elemental distribution using the real CBCT image was also predictable. CONCLUSIONS: This study demonstrated the potential of using computer vision for medical data preparation and analysis, which could have important implications for improving patient outcomes, especially in adaptive radiation therapy.

Radiomics analysis of [18F]-fluoro-2-deoxyglucose positron emission tomography for the prediction of cervical lymph node metastasis in tongue squamous cell carcinoma.

OBJECTIVES: This study aimed to create a predictive model for cervical lymph node metastasis (CLNM) in patients with tongue squamous cell carcinoma (SCC) based on radiomics features detected by [18F]-fluoro-2-deoxyglucose (18F-FDG) positron emission tomography (PET). METHODS: A total of 40 patients with tongue SCC who underwent 18F-FDG PET imaging during their first medical examination were enrolled. During the follow-up period (mean 28 months), 20 patients had CLNM, including six with late CLNM, whereas the remaining 20 patients did not have CLNM. Radiomics features were extracted from 18F-FDG PET images of all patients irrespective of metal artifact, and clinicopathological factors were obtained from the medical records. Late CLNM was defined as the CLNM that occurred after major treatment. The least absolute shrinkage and selection operator (LASSO) model was used for radiomics feature selection and sequential data fitting. The receiver operating characteristic curve analysis was used to assess the predictive performance of the 18F-FDG PET-based model and clinicopathological factors model (CFM) for CLNM. RESULTS: Six radiomics features were selected from LASSO analysis. The average values of the area under the curve (AUC), accuracy, sensitivity, and specificity of radiomics analysis for predicting CLNM from 18F-FDG PET images were 0.79, 0.68, 0.65, and 0.70, respectively. In contrast, those of the CFM were 0.54, 0.60, 0.60, and 0.60, respectively. The 18F-FDG PET-based model showed significantly higher AUC than that of the CFM. CONCLUSIONS: The 18F-FDG PET-based model has better potential for diagnosing CLNM and predicting late CLNM in patients with tongue SCC than the CFM.

Development of a more accurate Geant4 quantum molecular dynamics model for hadron therapy.

Objective. Although in heavy-ion therapy, the quantum molecular dynamics (QMD) model is one of the most fundamental physics models providing an accurate daughter-ion production yield in the final state, there are still non-negligible differences with the experimental results. The aim of this study is to improve fragment production in water phantoms by developing a more accurate QMD model in Geant4.Approach. A QMD model was developed by implementing modern Skyrme interaction parameter sets, as well as by incorporating with an ad hocα-cluster model in the initial nuclear state. Two adjusting parameters were selected that can significantly affect the fragment productions in the QMD model: the radius to discriminate a cluster to which nucleons belong after the nucleus-nucleus reaction, denoted byR, and the squared standard deviation of the Gaussian packet, denoted byL. Squared Mahalanobis's distance of fragment yields and angular distributions with 1, 2, and the higher atomic number for the produced fragments were employed as objective functions, and multi-objective optimization (MOO), which make it possible to compare quantitatively the simulated production yields with the reference experimental data, was performed.Main results. The MOO analysis showed that the QMD model with modern Skyrme parameters coupled with the proposedα-cluster model, denoted as SkM*α, can drastically improve light fragments yields in water. In addition, the proposed model reproduced the kinetic energy distribution of the fragments accurately. The optimizedLin SkM*αwas confirmed to be realistic by the charge radii analysis in the ground state formation.Significance. The proposed framework using MOO was demonstrated to be very useful in judging the superiority of the proposed nuclear model. The optimized QMD model is expected to improve the accuracy of heavy-ion therapy dosimetry.

Virtual computed-tomography system for deep-learning-based material decomposition.

Objective.Material decomposition (MD) evaluates the elemental composition of human tissues and organs via computed tomography (CT) and is indispensable in correlating anatomical images with functional ones. A major issue in MD is inaccurate elemental information about the real human body. To overcome this problem, we developed a virtual CT system model, by which various reconstructed images can be generated based on ICRP110 human phantoms with information about six major elements (H, C, N, O, P, and Ca).Approach.We generated CT datasets labelled with accurate elemental information using the proposed generative CT model and trained a deep learning (DL)-based model to estimate the material distribution with the ICRP110 based human phantom as well as the digital Shepp-Logan phantom. The accuracy in quad-, dual-, and single-energy CT cases was investigated. The influence of beam-hardening artefacts, noise, and spectrum variations were analysed with testing datasets including elemental density and anatomical shape variations.Main results.The results indicated that this DL approach can realise precise MD, even with single-energy CT images. Moreover, noise, beam-hardening artefacts, and spectrum variations were shown to have minimal impact on the MD.Significance.Present results suggest that the difficulty to prepare a large CT database can be solved by introducing the virtual CT system and the proposed technique can be applied to clinical radiodiagnosis and radiotherapy.

Training of deep cross-modality conversion models with a small data set, and their application in megavoltage CT to kilovoltage CT conversion.

PURPOSE: In recent years, deep learning-based image processing has emerged as a valuable tool for medical imaging owing to its high performance. However, the quality of deep learning-based methods heavily relies on the amount of training data; the high cost of acquiring a large data set is a limitation to their utilization in medical fields. Herein, based on deep learning, we developed a computed tomography (CT) modality conversion method requiring only a few unsupervised images. METHODS: The proposed method is based on cycle-consistency generative adversarial network (CycleGAN) with several extensions tailored for CT images, which aims at preserving the structure in the processed images and reducing the amount of training data. This method was applied to realize the conversion of megavoltage computed tomography (MVCT) to kilovoltage computed tomography (kVCT) images. Training was conducted using several data sets acquired from patients with head and neck cancer. The size of the data sets ranged from 16 slices (two patients) to 2745 slices (137 patients) for MVCT and 2824 slices (98 patients) for kVCT. RESULTS: The required size of the training data was found to be as small as a few hundred slices. By statistical and visual evaluations, the quality improvement and structure preservation of the MVCT images converted by the proposed model were investigated. As a clinical benefit, it was observed by medical doctors that the converted images enhanced the precision of contouring. CONCLUSIONS: We developed an MVCT to kVCT conversion model based on deep learning, which can be trained using only a few hundred unpaired images. The stability of the model against changes in data size was demonstrated. This study promotes the reliable use of deep learning in clinical medicine by partially answering commonly asked questions, such as "Is our data sufficient?" and "How much data should we acquire?"

Prediction of out-of-field recurrence after chemoradiotherapy for cervical cancer using a combination model of clinical parameters and magnetic resonance imaging radiomics: a multi-institutional study of the Japanese Radiation Oncology Study Group.

We retrospectively assessed whether magnetic resonance imaging (MRI) radiomics combined with clinical parameters can improve the predictability of out-of-field recurrence (OFR) of cervical cancer after chemoradiotherapy. The data set was collected from 204 patients with stage IIB (FIGO: International Federation of Gynecology and Obstetrics 2008) cervical cancer who underwent chemoradiotherapy at 14 Japanese institutes. Of these, 180 patients were finally included for analysis. OFR-free survival was calculated using the Kaplan-Meier method, and the statistical significance of clinicopathological parameters for the OFR-free survival was evaluated using the log-rank test and Cox proportional-hazards model. Prediction of OFR from the analysis of diffusion-weighted images (DWI) and T2-weighted images of pretreatment MRI was done using the least absolute shrinkage and selection operator (LASSO) model for engineering image feature extraction. The accuracy of prediction was evaluated by 5-fold cross-validation of the receiver operating characteristic (ROC) analysis. Para-aortic lymph node metastasis (p = 0.003) was a significant prognostic factor in univariate and multivariate analyses. ROC analysis showed an area under the curve (AUC) of 0.709 in predicting OFR using the pretreatment status of para-aortic lymph node metastasis, 0.667 using the LASSO model for DWIs and 0.602 using T2 weighted images. The AUC improved to 0.734 upon combining the pretreatment status of para-aortic lymph node metastasis with that from the LASSO model for DWIs. Combining MRI radiomics with clinical parameters improved the accuracy of predicting OFR after chemoradiotherapy for locally advanced cervical cancer.

Automatic contour segmentation of cervical cancer using artificial intelligence.

In cervical cancer treatment, radiation therapy is selected based on the degree of tumor progression, and radiation oncologists are required to delineate tumor contours. To reduce the burden on radiation oncologists, an automatic segmentation of the tumor contours would prove useful. To the best of our knowledge, automatic tumor contour segmentation has rarely been applied to cervical cancer treatment. In this study, diffusion-weighted images (DWI) of 98 patients with cervical cancer were acquired. We trained an automatic tumor contour segmentation model using 2D U-Net and 3D U-Net to investigate the possibility of applying such a model to clinical practice. A total of 98 cases were employed for the training, and they were then predicted by swapping the training and test images. To predict tumor contours, six prediction images were obtained after six training sessions for one case. The six images were then summed and binarized to output a final image through automatic contour segmentation. For the evaluation, the Dice similarity coefficient (DSC) and Hausdorff distance (HD) was applied to analyze the difference between tumor contour delineation by radiation oncologists and the output image. The DSC ranged from 0.13 to 0.93 (median 0.83, mean 0.77). The cases with DSC <0.65 included tumors with a maximum diameter < 40 mm and heterogeneous intracavitary concentration due to necrosis. The HD ranged from 2.7 to 9.6 mm (median 4.7 mm). Thus, the study confirmed that the tumor contours of cervical cancer can be automatically segmented with high accuracy.

kV-kV and kV-MV DECT based estimation of proton stopping power ratio - a simulation study.

PURPOSE: This study aims to estimate the proton stopping power ratio (SPR) by using 80-120 kV and 120 kV-6 MV dual-energy CT (DECT) in a fully simulation-based approach for proton therapy dose calculations. METHODS: Based on a virtual CT system, a two-step approach is applied to obtain the reference attenuation coefficient for image reconstruction. The effective atomic number (EAN) and electron density ratio (EDR) are estimated from two CT scans. The SPR is estimated using a calibration based on known materials to obtain a piecewise linear relationship between the EAN and the logarithm of the mean excitation energy, lnIm. The calibration phantoms are constructed from reference tissue materials taken from ICRU Report 44. Our approach is evaluated through using the ICRP110 human phantom. The respective influences of noise and beam hardening effects are studied. RESULTS: With the beam hardening correction applied, the results of 120 kV-6 MV DECT are comparable to those of 80-120 kV DECT in predicting the EAN, but the former demonstrated a clear improvement in predicting the EDR and SPR. The 120 kV-6 MV DECT is able to predict the SPR within an accuracy of 10% for lung tissue and 5% for pelvis tissue, thereby outperforming the 80-120 kV DECT. CONCLUSIONS: The 120 kV-6 MV DECT is less sensitive to noise but more susceptible to beam hardening effects. By applying beam hardening correction, the 120 kV-6 MV DECT can predict the SPR more accurately than the 80-120 kV DECT. To utilize our DECT approach most effectively, high-quality reconstructed images are required.

Computed tomography image representation using the Legendre polynomial and spherical harmonics functions.

The representation of computed tomography (CT) images using the Legendre polynomial (LPF) and spherical harmonics (SHF) functions was investigated. We selected 100 two-dimensional (2D) CT images of 10 lung cancer patients and 33 three-dimensional (3D) CT images of head and neck cancer patients. The reproducibility of these special functions was evaluated in terms of the normalized cross-correlation (NCC). For the 2D images, the NCC was 0.990 ± 0.002 (1sd) with an LPF of order 70, whereas for the 3D images, the NCC was 0.971 ± 0.004 (1sd) with an SHF of degree 70. The results showed that the LPF was more efficient than the Fourier series. As the thoracic and head areas are cylindrical and spherical, respectively, expansions with the LPF and SHF achieved an efficient representation of the human body. CT image representation with analytical functions can be potentially beneficial, such as in X-ray scattering estimation.

Retrospective dose reconstruction of prostate stereotactic body radiotherapy using cone-beam CT and a log file during VMAT delivery with flattening-filter-free mode.

This study aimed to reconstruct the dose distribution of single fraction of stereotactic body radiotherapy for patients with prostate cancer using cone-beam computed tomography (CBCT) and a log file during volumetric-modulated arc therapy (VMAT) delivery with flattening-filter-free (FFF) mode. Twenty patients with clinically localized prostate cancer were treated with FFF-VMAT, and projection images for in-treatment CBCT (iCBCT) imaging were concomitantly acquired with a log file. A D95 dose of 36.25 Gy in five fractions was prescribed to each planning target volume (PTV) on each treatment planning CT (pCT). Deformed pCT (dCT) was obtained from the iCBCT using a hybrid deformable image registration algorithm. Dose distributions on the dCT were calculated using Pinnacle3 v9.10 by converting the log file data to Pinnacle3 data format using an in-house software. Dose warping was performed by referring to deformation vector fields calculated from pCT and dCT. Reconstructed dose distribution was compared with that of the original plan. Dose differences between the original and reconstructed dose distributions were within 3% at the isocenter and observed in PTV and organ-at-risk (OAR) regions. Differences in OAR regions were relatively larger than those in the PTV, presumably because OARs were more deformed than the PTV. Therefore, our method can be used successfully to reconstruct the dose distributions of one fraction using iCBCT and a log file during FFF-VMAT delivery.

Publisher Correction: Prediction of malignant glioma grades using contrast-enhanced T1-weighted and T2-weighted magnetic resonance images based on a radiomic analysis.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Acceptable fetal dose using flattening filter-free volumetric arc therapy (FFF VMAT) in postoperative chemoradiotherapy of tongue cancer during pregnancy.

Optimizing irradiation protocols for pregnant women is challenging, because there are few cases and a dearth of fetal dosimetry data. We cared for a 36-year-old pregnant woman with tongue cancer. Prior to treatment, we compared three intensity-modulated radiation therapy (IMRT) techniques, including helical tomotherapy, volumetric arc therapy (VMAT), and flattening-filter free VMAT (FFF-VMAT) using treatment planning software. FFF-VMAT achieved the minimum fetal exposure and was selected as the optimal modality. We prescribed 66 Gy to the involved nodes, 60 Gy to the tumor bed and ipsilateral neck, and 54 Gy to the contralateral neck over 33 fractions. To confirm the out-of-field exposure per fraction, surface doses and the rectal dose were measured during FFF-VMAT delivery. Postoperative chemoradiotherapy was delivered using IMRT and a cisplatin regimen. Without any shielding, the total fetal dose was 0.03 Gy, within the limits established by the ICRP. A healthy girl was born vaginally at 37 weeks' gestation.

Prediction of malignant glioma grades using contrast-enhanced T1-weighted and T2-weighted magnetic resonance images based on a radiomic analysis.

We conducted a feasibility study to predict malignant glioma grades via radiomic analysis using contrast-enhanced T1-weighted magnetic resonance images (CE-T1WIs) and T2-weighted magnetic resonance images (T2WIs). We proposed a framework and applied it to CE-T1WIs and T2WIs (with tumor region data) acquired preoperatively from 157 patients with malignant glioma (grade III: 55, grade IV: 102) as the primary dataset and 67 patients with malignant glioma (grade III: 22, grade IV: 45) as the validation dataset. Radiomic features such as size/shape, intensity, histogram, and texture features were extracted from the tumor regions on the CE-T1WIs and T2WIs. The Wilcoxon-Mann-Whitney (WMW) test and least absolute shrinkage and selection operator logistic regression (LASSO-LR) were employed to select the radiomic features. Various machine learning (ML) algorithms were used to construct prediction models for the malignant glioma grades using the selected radiomic features. Leave-one-out cross-validation (LOOCV) was implemented to evaluate the performance of the prediction models in the primary dataset. The selected radiomic features for all folds in the LOOCV of the primary dataset were used to perform an independent validation. As evaluation indices, accuracies, sensitivities, specificities, and values for the area under receiver operating characteristic curve (or simply the area under the curve (AUC)) for all prediction models were calculated. The mean AUC value for all prediction models constructed by the ML algorithms in the LOOCV of the primary dataset was 0.902 ± 0.024 (95% CI (confidence interval), 0.873-0.932). In the independent validation, the mean AUC value for all prediction models was 0.747 ± 0.034 (95% CI, 0.705-0.790). The results of this study suggest that the malignant glioma grades could be sufficiently and easily predicted by preparing the CE-T1WIs, T2WIs, and tumor delineations for each patient. Our proposed framework may be an effective tool for preoperatively grading malignant gliomas.

Optimization of treatment strategy by using a machine learning model to predict survival time of patients with malignant glioma after radiotherapy.

The purpose of this study was to predict the survival time of patients with malignant glioma after radiotherapy with high accuracy by considering additional clinical factors and optimize the prescription dose and treatment duration for individual patient by using a machine learning model. A total of 35 patients with malignant glioma were included in this study. The candidate features included 12 clinical features and 192 dose-volume histogram (DVH) features. The appropriate input features and parameters of the support vector machine (SVM) were selected using the genetic algorithm based on Akaike's information criterion, i.e. clinical, DVH, and both clinical and DVH features. The prediction accuracy of the SVM models was evaluated through a leave-one-out cross-validation test with residual error, which was defined as the absolute difference between the actual and predicted survival times after radiotherapy. Moreover, the influences of various values of prescription dose and treatment duration on the predicted survival time were evaluated. The prediction accuracy was significantly improved with the combined use of clinical and DVH features compared with the separate use of both features (P < 0.01, Wilcoxon signed rank test). Mean ± standard deviation of the leave-one-out cross-validation using the combined clinical and DVH features, only clinical features and only DVH features were 104.7 ± 96.5, 144.2 ± 126.1 and 204.5 ± 186.0 days, respectively. The prediction accuracy could be improved with the combination of clinical and DVH features, and our results show the potential to optimize the treatment strategy for individual patients based on a machine learning model.

Improvement of the robustness to set up error by a virtual bolus in total scalp irradiation with Helical TomoTherapy.

Intensity-modulated radiation therapy has recently been used for total scalp irradiation. In inverse planning, the treatment planning system increases the fluence of tangential beam near the skin surface to counter the build-up region. Consequently, the dose to the skin surface increases even with small setup errors. Replacing the electron density of the surrounding air of some thickness with a virtual bolus during optimization could suppress the extremely high fluence near the skin. We confirmed the usefulness of a virtual bolus in total scalp irradiation. For each patient, two beams were planned, one with and the other without a virtual bolus. The dose distribution was calculated using computed tomography images that were shifted to simulate setup errors. The hot spot dose was suppressed in the plans using a virtual bolus. In conclusion, using a virtual bolus improved the robustness to setup errors.

Radiomics Analysis for Glioma Malignancy Evaluation Using Diffusion Kurtosis and Tensor Imaging.

PURPOSE: A noninvasive diagnostic method to predict the degree of malignancy accurately would be of great help in glioma management. This study aimed to create a highly accurate machine learning model to perform glioma grading. METHODS AND MATERIALS: Preoperative magnetic resonance imaging acquired for cases of glioma operated on at our institution from October 2014 through January 2018 were obtained retrospectively. Six types of magnetic resonance imaging sequences (T2-weighted image, diffusion-weighted image, apparent diffusion coefficient [ADC], fractional anisotropy, and mean kurtosis [MK]) were chosen for analysis; 476 features were extracted semiautomatically for each sequence (2856 features in total). Recursive feature elimination was used to select significant features for a machine learning model that distinguishes glioblastoma from lower-grade glioma (grades 2 and 3). RESULTS: Fifty-five data sets from 54 cases were obtained (14 grade 2 gliomas, 12 grade 3 gliomas, and 29 glioblastomas), of which 44 and 11 data sets were used for machine learning and independent testing, respectively. We detected 504 features with significant differences (false discovery rate <0.05) between glioblastoma and lower-grade glioma. The most accurate machine learning model was created using 6 features extracted from the ADC and MK images. In the logistic regression, the area under the curve was 0.90 ± 0.05, and the accuracy of the test data set was 0.91 (10 out of 11); using a support vector machine, they were 0.93 ± 0.03 and 0.91 (10 out of 11), respectively (kernel, radial basis function; c = 1.0). CONCLUSIONS: Our machine learning model accurately predicted glioma tumor grade. The ADC and MK sequences produced particularly useful features.

Standardization of imaging features for radiomics analysis.

Radiomics has the potential to provide tumor characteristics with noninvasive and repeatable way. The purpose of this paper is to evaluate the standardization effect of imaging features for radiomics analysis. For this purpose, we prepared two CT databases ; one includes 40 non-small cell lung cancer (NSCLC) patients for whom tumor biopsies was performed before stereotactic body radiation therapy in The University of Tokyo Hospital, and the other includes 29 early-stage NSCLC datasets from the Cancer Imaging Archive. The former was used as the training data, whereas the later was used as the test data in the evaluation of the prediction model. In total, 476 imaging features were extracted from each data. Then, both training and test data were standardized as the min-max normalization, the z-score normalization, and the whitening from the principle component analysis. All of standardization strategies improved the accuracy for the histology prediction. The area under the receiver observed characteristics curve was 0.725, 0.789, and 0.785 in above standardizations, respectively. Radiomics analysis has shown that robust features have a high prognostic power in predicting early-stage NSCLC histology subtypes. The performance was able to be improved by standardizing the data in the feature space. J. Med. Invest. 66 : 35-37, February, 2019.