Model Generalizability Investigation for GFCE-MRI Synthesis in NPC Radiotherapy Using Multi-Institutional Patient-Based Data Normalization.

Recently, deep learning has been demonstrated to be feasible in eliminating the use of gadoliniumbased contrast agents (GBCAs) through synthesizing gadolinium-free contrast-enhanced MRI (GFCE-MRI) from contrast-free MRI sequences, providing the community with an alternative to get rid of GBCAs-associated safety issues in patients. Nevertheless, generalizability assessment of the GFCE-MRI model has been largely challenged by the high inter-institutional heterogeneity of MRI data, on top of the scarcity of multi-institutional data itself. Although various data normalization methods have been adopted to address the heterogeneity issue, it has been limited to single-institutional investigation and there is no standard normalization approach presently. In this study, we aimed at investigating generalizability of GFCE-MRI model using data from seven institutions by manipulating heterogeneity of MRI data under five popular normalization approaches. Three state-of-the-art neural networks were applied to map from T1-weighted and T2-weighted MRI to contrast-enhanced MRI (CE-MRI) for GFCE-MRI synthesis in patients with nasopharyngeal carcinoma. MRI data from three institutions were used separately to generate three uni-institution models and jointly for a tri-institution model. The five normalization methods were applied to normalize the data of each model. MRI data from the remaining four institutions served as external cohorts for model generalizability assessment. Quality of GFCE-MRI was quantitatively evaluated against ground-truth CE-MRI using mean absolute error (MAE) and peak signal-to-noise ratio(PSNR). Results showed that performance of all uni-institution models remarkably dropped on the external cohorts. By contrast, model trained using multi-institutional data with Z-Score normalization yielded the best model generalizability improvement.

Multi-omics fusion with soft labeling for enhanced prediction of distant metastasis in nasopharyngeal carcinoma patients after radiotherapy.

Omics fusion has emerged as a crucial preprocessing approach in medical image processing, significantly assisting several studies. One of the challenges encountered in integrating omics data is the unpredictability arising from disparities in data sources and medical imaging equipment. Due to these differences, the distribution of omics futures exhibits spatial heterogeneity, diminishing their capacity to enhance subsequent tasks. To overcome this challenge and facilitate the integration of their joint application to specific medical objectives, this study aims to develop a fusion methodology for nasopharyngeal carcinoma (NPC) distant metastasis prediction to mitigate the disparities inherent in omics data. The multi-kernel late-fusion method can reduce the impact of these differences by mapping the features using the most suiTable single-kernel function and then combining them in a high-dimensional space that can effectively represent the data. The proposed approach in this study employs a distinctive framework incorporating a label-softening technique alongside a multi-kernel-based Radial basis function (RBF) neural network to address these limitations. An efficient representation of the data may be achieved by utilizing the multi-kernel to map the inherent features and then merging them in a space with many dimensions. However, the inflexibility of label fitting poses a constraint on using multi-kernel late-fusion methods in complex NPC datasets, hence affecting the efficacy of general classifiers in dealing with high-dimensional characteristics. The label softening increases the disparity between the two cohorts, providing a more flexible structure for allocating labels. The proposed model is evaluated on multi-omics datasets, and the results demonstrate its strength and effectiveness in predicting distant metastasis of NPC patients.

Computed Tomography-Based Radiomics for Long-Term Prognostication of High-Risk Localized Prostate Cancer Patients Received Whole Pelvic Radiotherapy.

Given the high death rate caused by high-risk prostate cancer (PCa) (>40%) and the reliability issues associated with traditional prognostic markers, the purpose of this study is to investigate planning computed tomography (pCT)-based radiomics for the long-term prognostication of high-risk localized PCa patients who received whole pelvic radiotherapy (WPRT). This is a retrospective study with methods based on best practice procedures for radiomics research. Sixty-four patients were selected and randomly assigned to training (n = 45) and testing (n = 19) cohorts for radiomics model development with five major steps: pCT image acquisition using a Philips Big Bore CT simulator; multiple manual segmentations of clinical target volume for the prostate (CTVprostate) on the pCT images; feature extraction from the CTVprostate using PyRadiomics; feature selection for overfitting avoidance; and model development with three-fold cross-validation. The radiomics model and signature performances were evaluated based on the area under the receiver operating characteristic curve (AUC) as well as accuracy, sensitivity and specificity. This study's results show that our pCT-based radiomics model was able to predict the six-year progression-free survival of the high-risk localized PCa patients who received the WPRT with highly consistent performances (mean AUC: 0.76 (training) and 0.71 (testing)). These are comparable to findings of other similar studies including those using magnetic resonance imaging (MRI)-based radiomics. The accuracy, sensitivity and specificity of our radiomics signature that consisted of two texture features were 0.778, 0.833 and 0.556 (training) and 0.842, 0.867 and 0.750 (testing), respectively. Since CT is more readily available than MRI and is the standard-of-care modality for PCa WPRT planning, pCT-based radiomics could be used as a routine non-invasive approach to the prognostic prediction of WPRT treatment outcomes in high-risk localized PCa.

Comparing effectiveness of image perturbation and test retest imaging in improving radiomic model reliability.

Image perturbation is a promising technique to assess radiomic feature repeatability, but whether it can achieve the same effect as test-retest imaging on model reliability is unknown. This study aimed to compare radiomic model reliability based on repeatable features determined by the two methods using four different classifiers. A 191-patient public breast cancer dataset with 71 test-retest scans was used with pre-determined 117 training and 74 testing samples. We collected apparent diffusion coefficient images and manual tumor segmentations for radiomic feature extraction. Random translations, rotations, and contour randomizations were performed on the training images, and intra-class correlation coefficient (ICC) was used to filter high repeatable features. We evaluated model reliability in both internal generalizability and robustness, which were quantified by training and testing AUC and prediction ICC. Higher testing performance was found at higher feature ICC thresholds, but it dropped significantly at ICC = 0.95 for the test-retest model. Similar optimal reliability can be achieved with testing AUC = 0.7-0.8 and prediction ICC > 0.9 at the ICC threshold of 0.9. It is recommended to include feature repeatability analysis using image perturbation in any radiomic study when test-retest is not feasible, but care should be taken when deciding the optimal feature repeatability criteria.

Explainable machine learning via intra-tumoral radiomics feature mapping for patient stratification in adjuvant chemotherapy for locoregionally advanced nasopharyngeal carcinoma.

PURPOSE: This study aimed to discover intra-tumor heterogeneity signature and validate its predictive value for adjuvant chemotherapy (ACT) following concurrent chemoradiotherapy (CCRT) in locoregionally advanced nasopharyngeal carcinoma (LA-NPC). MATERIALS AND METHODS: 397 LA-NPC patients were retrospectively enrolled. Pre-treatment contrast-enhanced T1-weighted (CET1-w) MR images, clinical variables, and follow-up were retrospectively collected. We identified single predictive radiomic feature from primary gross tumor volume (GTVnp) and defined predicted subvolume by calculating voxel-wised feature mapping and within GTVnp. We independently validate predictive value of identified feature and associated predicted subvolume. RESULTS: Only one radiomic feature, gldm_DependenceVariance in 3 mm-sigma LoG-filtered image, was discovered as a signature. In the high-risk group determined by the signature, patients received CCRT + ACT achieved 3-year disease free survival (DFS) rate of 90% versus 57% (HR, 0.20; 95%CI, 0.05-0.94; P = 0.007) for CCRT alone. The multivariate analysis showed patients receiving CCRT + ACT had a HR of 0.21 (95%CI: 0.06-0.68, P = 0.009) for DFS compared to those receiving CCRT alone. The predictive value can also be generalized to the subvolume with multivariate HR of 0.27 (P = 0.017) for DFS. CONCLUSION: The signature with its heterogeneity mapping could be a reliable and explainable ACT decision-making tool in clinical practice.

Artificial intelligence-driven radiomics study in cancer: the role of feature engineering and modeling.

Modern medicine is reliant on various medical imaging technologies for non-invasively observing patients' anatomy. However, the interpretation of medical images can be highly subjective and dependent on the expertise of clinicians. Moreover, some potentially useful quantitative information in medical images, especially that which is not visible to the naked eye, is often ignored during clinical practice. In contrast, radiomics performs high-throughput feature extraction from medical images, which enables quantitative analysis of medical images and prediction of various clinical endpoints. Studies have reported that radiomics exhibits promising performance in diagnosis and predicting treatment responses and prognosis, demonstrating its potential to be a non-invasive auxiliary tool for personalized medicine. However, radiomics remains in a developmental phase as numerous technical challenges have yet to be solved, especially in feature engineering and statistical modeling. In this review, we introduce the current utility of radiomics by summarizing research on its application in the diagnosis, prognosis, and prediction of treatment responses in patients with cancer. We focus on machine learning approaches, for feature extraction and selection during feature engineering and for imbalanced datasets and multi-modality fusion during statistical modeling. Furthermore, we introduce the stability, reproducibility, and interpretability of features, and the generalizability and interpretability of models. Finally, we offer possible solutions to current challenges in radiomics research.

Integrating CT-based radiomic model with clinical features improves long-term prognostication in high-risk prostate cancer.

Objective: High-risk prostate cancer (PCa) is often treated by prostate-only radiotherapy (PORT) owing to its favourable toxicity profile compared to whole-pelvic radiotherapy. Unfortunately, more than 50% patients still developed disease progression following PORT. Conventional clinical factors may be unable to identify at-risk subgroups in the era of precision medicine. In this study, we aimed to investigate the prognostic value of pre-treatment planning computed tomography (pCT)-based radiomic features and clinical attributes to predict 5-year progression-free survival (PFS) in high-risk PCa patients following PORT. Materials and methods: A total of 176 biopsy-confirmed PCa patients who were treated at the Hong Kong Princess Margaret Hospital were retrospectively screened for eligibility. Clinical data and pCT of one hundred eligible high-risk PCa patients were analysed. Radiomic features were extracted from the gross-tumour-volume (GTV) with and without applying Laplacian-of-Gaussian (LoG) filter. The entire patient cohort was temporally stratified into a training and an independent validation cohort in a ratio of 3:1. Radiomics (R), clinical (C) and radiomic-clinical (RC) combined models were developed by Ridge regression through 5-fold cross-validation with 100 iterations on the training cohort. A model score was calculated for each model based on the included features. Model classification performance on 5-year PFS was evaluated in the independent validation cohort by average area-under-curve (AUC) of receiver-operating-characteristics (ROC) curve and precision-recall curve (PRC). Delong's test was used for model comparison. Results: The RC combined model which contains 6 predictive features (tumour flatness, root-mean-square on fine LoG-filtered image, prostate-specific antigen serum concentration, Gleason score, Roach score and GTV volume) was the best-performing model (AUC = 0.797, 95%CI = 0.768-0.826), which significantly outperformed the R-model (AUC = 0.795, 95%CI = 0.774-0.816) and C-model (AUC = 0.625, 95%CI = 0.585-0.665) in the independent validation cohort. Besides, only the RC model score significantly classified patients in both cohorts into progression and progression-free groups regarding their 5-year PFS (p< 0.05). Conclusion: Combining pCT-based radiomic and clinical attributes provided superior prognostication value regarding 5-year PFS in high-risk PCa patients following PORT. A large multi-centre study will potentially aid clinicians in implementing personalised treatment for this vulnerable subgroup in the future.

Multi-omics to predict acute radiation esophagitis in patients with lung cancer treated with intensity-modulated radiation therapy.

PURPOSE: The study aimed to predict acute radiation esophagitis (ARE) with grade ≥ 2 for patients with locally advanced lung cancer (LALC) treated with intensity-modulated radiation therapy (IMRT) using multi-omics features, including radiomics and dosiomics. METHODS: 161 patients with stage IIIA-IIIB LALC who received chemoradiotherapy (CRT) or radiotherapy by IMRT with a prescribed dose from 45 to 70 Gy from 2015 to 2019 were enrolled retrospectively. All the toxicity gradings were given following the Common Terminology Criteria for Adverse Events V4.0. Multi-omics features, including radiomics, dosiomics (including dose-volume histogram dosimetric parameters), were extracted based on the planning CT image and three-dimensional dose distribution. All data were randomly divided into training cohorts (N = 107) and testing cohorts (N = 54). In the training cohorts, features with reliably high outcome relevance and low redundancy were selected under random patient subsampling. Four classification models (using clinical factors (CF) only, using radiomics features (RFs) only, dosiomics features (DFs) only, and the hybrid features (HFs) containing clinical factors, radiomics and dosiomics) were constructed employing the Ridge classifier using two-thirds of randomly selected patients as the training cohort. The remaining patient was treated as the testing cohort. A series of models were built with 30 times training-testing splits. Their performances were assessed using the area under the ROC curve (AUC) and accuracy. RESULTS: Among all patients, 51 developed ARE grade ≥ 2, with an incidence of 31.7%. Next, 8990 radiomics and 213 dosiomics features were extracted, and 3, 6, 12, and 13 features remained after feature selection in the CF, DF, RF and DF models, respectively. The RF and HF models achieved similar classification performance, with the training and testing AUCs of 0.796 ± 0.023 (95% confidence interval (CI [0.79, 0.80])/0.744 ± 0.044 (95% CI [0.73, 0.76]) and 0.801 ± 0.022 (95% CI [0.79, 0.81]) (p = 0.74), respectively. The model performances using CF and DF features were poorer, with training and testing AUCs of 0.573 ± 0.026 (95% CI [0.56, 0.58])/ 0.509 ± 0.072 (95% CI [0.48, 0.53]) and 0.679 ± 0.027 (95% CI [0.67, 0.69])/0.604 ± 0.041 (95% CI [0.53, 0.63]) compared with the above two models (p < 0.001), respectively. CONCLUSIONS: In LALC patients treated with CRT IMRT, the ARE grade ≥ 2 can be predicted using the pretreatment radiotherapy image features. To predict ARE, the multi-omics features had similar predictability with radiomics features; however, the dosiomics features and clinical factors had a limited classification performance.

Association of Multi-Phasic MR-Based Radiomic and Dosimetric Features with Treatment Response in Unresectable Hepatocellular Carcinoma Patients following Novel Sequential TACE-SBRT-Immunotherapy.

This study aims to investigate the association of pre-treatment multi-phasic MR-based radiomics and dosimetric features with treatment response to a novel sequential trans-arterial chemoembolization (TACE) plus stereotactic body radiotherapy (SBRT) plus immunotherapy regimen in unresectable Hepatocellular Carcinoma (HCC) sub-population. Twenty-six patients with unresectable HCC were retrospectively analyzed. Radiomic features were extracted from 42 lesions on arterial phase (AP) and portal-venous phase (PVP) MR images. Delta-phase (DeltaP) radiomic features were calculated as AP-to-PVP ratio. Dosimetric data of the tumor was extracted from dose-volume-histograms. A two-sided independent Mann-Whitney U test was used to assess the clinical association of each feature, and the classification performance of each significant independent feature was assessed using logistic regression. For the 3-month timepoint, four DeltaP-derived radiomics that characterize the temporal change in intratumoral randomness and uniformity were the only contributors to the treatment response association (p-value = 0.038-0.063, AUC = 0.690-0.766). For the 6-month timepoint, DeltaP-derived radiomic features (n = 4) maintained strong clinical associations with the treatment response (p-value = 0.047-0.070, AUC = 0.699-0.788), additional AP-derived radiomic features (n = 4) that reflect baseline tumoral arterial-enhanced signal pattern and tumor morphology (n = 1) that denotes initial tumor burden were shown to have strong associations with treatment response (p-value = 0.028-0.074, AUC = 0.719-0.773). This pilot study successfully demonstrated associations of pre-treatment multi-phasic MR-based radiomics with tumor response to the novel treatment regimen.

Radiomic feature repeatability and its impact on prognostic model generalizability: A multi-institutional study on nasopharyngeal carcinoma patients.

BACKGROUND AND PURPOSE: To investigate the radiomic feature (RF) repeatability via perturbation and its impact on cross-institutional prognostic model generalizability in Nasopharyngeal Carcinoma (NPC) patients. MATERIALS AND METHODS: 286 and 183 NPC patients from two institutions were included for model training and validation. Perturbations with random translations and rotations were applied to contrast-enhanced T1-weighted (CET1-w) MR images. RFs were extracted from primary tumor volume under a wide range of image filtering and discretization settings. RF repeatability was assessed by intraclass correlation coefficient (ICC), which was used to equally separate the RFs into low- and high-repeatable groups by the median value. After feature selection, multivariate Cox regression and Kaplan-Meier analysis were independently employed to develop and analyze prognostic models. Concordance index (C-index) and P-value from log-rank test were used to assess model performance. RESULTS: Most textural RFs from high-pass wavelet-filtered images were susceptible to image perturbations. It was more prominent when a smaller discretization bin number was used (e.g., 8, mean ICC = 0.69). Using high-repeatable RFs for model development yielded a significantly higher C-index (0.63) in the validation cohort than when only low-repeatable RFs were used (0.57, P = 0.024), suggesting higher model generalizability. Besides, significant risk stratification in the validation cohort was observed only when high-repeatable RFs were used (P < 0.001). CONCLUSION: Repeatability of RFs from high-pass wavelet-filtered CET1-w MR images of primary NPC tumor was poor, particularly when a smaller bin number was used. Exclusive use of high-repeatable RFs is suggested to safeguard model generalizability for wide-spreading clinical utilization.

Radiomics-Based Detection of COVID-19 from Chest X-ray Using Interpretable Soft Label-Driven TSK Fuzzy Classifier.

The COVID-19 pandemic has posed a significant global public health threat with an escalating number of new cases and death toll daily. The early detection of COVID-related CXR abnormality potentially allows the early isolation of suspected cases. Chest X-Ray (CXR) is a fast and highly accessible imaging modality. Recently, a number of CXR-based AI models have been developed for the automated detection of COVID-19. However, most existing models are difficult to interpret due to the use of incomprehensible deep features in their models. Confronted with this, we developed an interpretable TSK fuzzy system in this study for COVID-19 detection using radiomics features extracted from CXR images. There are two main contributions. (1) When TSK fuzzy systems are applied to classification tasks, the commonly used binary label matrix of training samples is transformed into a soft one in order to learn a more discriminant transformation matrix and hence improve classification accuracy. (2) Based on the assumption that the samples in the same class should be kept as close as possible when they are transformed into the label space, the compactness class graph is introduced to avoid overfitting caused by label matrix relaxation. Our proposed model for a multi-categorical classification task (COVID-19 vs. No-Findings vs. Pneumonia) was evaluated using 600 CXR images from publicly available datasets and compared against five state-of-the-art AI models in aspects of classification accuracy. Experimental findings showed that our model achieved classification accuracy of over 83%, which is better than the state-of-the-art models, while maintaining high interpretability.

Lung Subregion Partitioning by Incremental Dose Intervals Improves Omics-Based Prediction for Acute Radiation Pneumonitis in Non-Small-Cell Lung Cancer Patients.

Purpose: To evaluate the effectiveness of features obtained from our proposed incremental-dose-interval-based lung subregion segmentation (IDLSS) for predicting grade ≥ 2 acute radiation pneumonitis (ARP) in lung cancer patients upon intensity-modulated radiotherapy (IMRT). (1) Materials and Methods: A total of 126 non-small-cell lung cancer patients treated with IMRT were retrospectively analyzed. Five lung subregions (SRs) were generated by the intersection of the whole lung (WL) and five sub-regions receiving incremental dose intervals. A total of 4610 radiomics features (RF) from pre-treatment planning computed tomographic (CT) and 213 dosiomics features (DF) were extracted. Six feature groups, including WL-RF, WL-DF, SR-RF, SR-DF, and the combined feature sets of WL-RDF and SR-RDF, were generated. Features were selected by using a variance threshold, followed by a Student t-test. Pearson's correlation test was applied to remove redundant features. Subsequently, Ridge regression was adopted to develop six models for ARP using the six feature groups. Thirty iterations of resampling were implemented to assess overall model performance by using the area under the Receiver-Operating-Characteristic curve (AUC), accuracy, precision, recall, and F1-score. (2) Results: The SR-RDF model achieved the best classification performance and provided significantly better predictability than the WL-RDF model in training cohort (Average AUC: 0.98 ± 0.01 vs. 0.90 ± 0.02, p < 0.001) and testing cohort (Average AUC: 0.88 ± 0.05 vs. 0.80 ± 0.04, p < 0.001). Similarly, predictability of the SR-DF model was significantly stronger than that of the WL-DF model in training cohort (Average AUC: 0.88 ± 0.03 vs. 0.70 ± 0.030, p < 0.001) and in testing cohort (Average AUC: 0.74 ± 0.08 vs. 0.65 ± 0.06, p < 0.001). By contrast, the SR-RF model significantly outperformed the WL-RF model only in the training set (Average AUC: 0.93 ± 0.02 vs. 0.85 ± 0.03, p < 0.001), but not in the testing set (Average AUC: 0.79 ± 0.05 vs. 0.77 ± 0.07, p = 0.13). (3) Conclusions: Our results demonstrated that the IDLSS method improved model performance for classifying ARP with grade ≥ 2 when using dosiomics or combined radiomics-dosiomics features.

Function-Wise Dual-Omics analysis for radiation pneumonitis prediction in lung cancer patients.

Purpose: This study investigates the impact of lung function on radiation pneumonitis prediction using a dual-omics analysis method. Methods: We retrospectively collected data of 126 stage III lung cancer patients treated with chemo-radiotherapy using intensity-modulated radiotherapy, including pre-treatment planning CT images, radiotherapy dose distribution, and contours of organs and structures. Lung perfusion functional images were generated using a previously developed deep learning method. The whole lung (WL) volume was divided into function-wise lung (FWL) regions based on the lung perfusion functional images. A total of 5,474 radiomics features and 213 dose features (including dosiomics features and dose-volume histogram factors) were extracted from the FWL and WL regions, respectively. The radiomics features (R), dose features (D), and combined dual-omics features (RD) were used for the analysis in each lung region of WL and FWL, labeled as WL-R, WL-D, WL-RD, FWL-R, FWL-D, and FWL-RD. The feature selection was carried out using ANOVA, followed by a statistical F-test and Pearson correlation test. Thirty times train-test splits were used to evaluate the predictability of each group. The overall average area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and f1-score were calculated to assess the performance of each group. Results: The FWL-RD achieved a significantly higher average AUC than the WL-RD group in the training (FWL-RD: 0.927 ± 0.031, WL-RD: 0.849 ± 0.064) and testing cohorts (FWL-RD: 0.885 ± 0.028, WL-RD: 0.762 ± 0.053, p < 0.001). When using radiomics features only, the FWL-R group yielded a better classification result than the model trained with WL-R features in the training (FWL-R: 0.919 ± 0.036, WL-R: 0.820 ± 0.052) and testing cohorts (FWL-R: 0.862 ± 0.028, WL-R: 0.750 ± 0.057, p < 0.001). The FWL-D group obtained an average AUC of 0.782 ± 0.032, obtaining a better classification performance than the WL-D feature-based model of 0.740 ± 0.028 in the training cohort, while no significant difference was observed in the testing cohort (FWL-D: 0.725 ± 0.064, WL-D: 0.710 ± 0.068, p = 0.54). Conclusion: The dual-omics features from different lung functional regions can improve the prediction of radiation pneumonitis for lung cancer patients under IMRT treatment. This function-wise dual-omics analysis method holds great promise to improve the prediction of radiation pneumonitis for lung cancer patients.

Improving radiomic model reliability using robust features from perturbations for head-and-neck carcinoma.

Background: Using high robust radiomic features in modeling is recommended, yet its impact on radiomic model is unclear. This study evaluated the radiomic model's robustness and generalizability after screening out low-robust features before radiomic modeling. The results were validated with four datasets and two clinically relevant tasks. Materials and methods: A total of 1,419 head-and-neck cancer patients' computed tomography images, gross tumor volume segmentation, and clinically relevant outcomes (distant metastasis and local-regional recurrence) were collected from four publicly available datasets. The perturbation method was implemented to simulate images, and the radiomic feature robustness was quantified using intra-class correlation of coefficient (ICC). Three radiomic models were built using all features (ICC > 0), good-robust features (ICC > 0.75), and excellent-robust features (ICC > 0.95), respectively. A filter-based feature selection and Ridge classification method were used to construct the radiomic models. Model performance was assessed with both robustness and generalizability. The robustness of the model was evaluated by the ICC, and the generalizability of the model was quantified by the train-test difference of Area Under the Receiver Operating Characteristic Curve (AUC). Results: The average model robustness ICC improved significantly from 0.65 to 0.78 (P< 0.0001) using good-robust features and to 0.91 (P< 0.0001) using excellent-robust features. Model generalizability also showed a substantial increase, as a closer gap between training and testing AUC was observed where the mean train-test AUC difference was reduced from 0.21 to 0.18 (P< 0.001) in good-robust features and to 0.12 (P< 0.0001) in excellent-robust features. Furthermore, good-robust features yielded the best average AUC in the unseen datasets of 0.58 (P< 0.001) over four datasets and clinical outcomes. Conclusions: Including robust only features in radiomic modeling significantly improves model robustness and generalizability in unseen datasets. Yet, the robustness of radiomic model has to be verified despite building with robust radiomic features, and tightly restricted feature robustness may prevent the optimal model performance in the unseen dataset as it may lower the discrimination power of the model.

Corrigendum: Investigation of a novel deep learning-based computed tomography perfusion mapping framework for functional lung avoidance radiotherapy.

[This corrects the article DOI: 10.3389/fonc.2021.644703.].

A Transfer Learning Framework for Deep Learning-Based CT-to-Perfusion Mapping on Lung Cancer Patients.

Purpose: Deep learning model has shown the feasibility of providing spatial lung perfusion information based on CT images. However, the performance of this method on lung cancer patients is yet to be investigated. This study aims to develop a transfer learning framework to evaluate the deep learning based CT-to-perfusion mapping method specifically on lung cancer patients. Methods: SPECT/CT perfusion scans of 33 lung cancer patients and 137 non-cancer patients were retrospectively collected from two hospitals. To adapt the deep learning model on lung cancer patients, a transfer learning framework was developed to utilize the features learned from the non-cancer patients. These images were processed to extract features from three-dimensional CT images and synthesize the corresponding CT-based perfusion images. A pre-trained model was first developed using a dataset of patients with lung diseases other than lung cancer, and subsequently fine-tuned specifically on lung cancer patients under three-fold cross-validation. A multi-level evaluation was performed between the CT-based perfusion images and ground-truth SPECT perfusion images in aspects of voxel-wise correlation using Spearman's correlation coefficient (R), function-wise similarity using Dice Similarity Coefficient (DSC), and lobe-wise agreement using mean perfusion value for each lobe of the lungs. Results: The fine-tuned model yielded a high voxel-wise correlation (0.8142 ± 0.0669) and outperformed the pre-trained model by approximately 8%. Evaluation of function-wise similarity indicated an average DSC value of 0.8112 ± 0.0484 (range: 0.6460-0.8984) for high-functional lungs and 0.8137 ± 0.0414 (range: 0.6743-0.8902) for low-functional lungs. Among the 33 lung cancer patients, high DSC values of greater than 0.7 were achieved for high functional volumes in 32 patients and low functional volumes in all patients. The correlations of the mean perfusion value on the left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe were 0.7314, 0.7134, 0.5108, 0.4765, and 0.7618, respectively. Conclusion: For lung cancer patients, the CT-based perfusion images synthesized by the transfer learning framework indicated a strong voxel-wise correlation and function-wise similarity with the SPECT perfusion images. This suggests the great potential of the deep learning method in providing regional-based functional information for functional lung avoidance radiation therapy.

Building reliable radiomic models using image perturbation.

Radiomic model reliability is a central premise for its clinical translation. Presently, it is assessed using test-retest or external data, which, unfortunately, is often scarce in reality. Therefore, we aimed to develop a novel image perturbation-based method (IPBM) for the first of its kind toward building a reliable radiomic model. We first developed a radiomic prognostic model for head-and-neck cancer patients on a training (70%) and evaluated on a testing (30%) cohort using C-index. Subsequently, we applied the IPBM to CT images of both cohorts (Perturbed-Train and Perturbed-Test cohort) to generate 60 additional samples for both cohorts. Model reliability was assessed using intra-class correlation coefficient (ICC) to quantify consistency of the C-index among the 60 samples in the Perturbed-Train and Perturbed-Test cohorts. Besides, we re-trained the radiomic model using reliable RFs exclusively (ICC > 0.75) to validate the IPBM. Results showed moderate model reliability in Perturbed-Train (ICC: 0.565, 95%CI 0.518-0.615) and Perturbed-Test (ICC: 0.596, 95%CI 0.527-0.670) cohorts. An enhanced reliability of the re-trained model was observed in Perturbed-Train (ICC: 0.782, 95%CI 0.759-0.815) and Perturbed-Test (ICC: 0.825, 95%CI 0.782-0.867) cohorts, indicating validity of the IPBM. To conclude, we demonstrated capability of the IPBM toward building reliable radiomic models, providing community with a novel model reliability assessment strategy prior to prospective evaluation.

Evaluation of Multisource Adaptive MRI Fusion for Gross Tumor Volume Delineation of Hepatocellular Carcinoma.

Purpose: Tumor delineation plays a critical role in radiotherapy for hepatocellular carcinoma (HCC) patients. The incorporation of MRI might improve the ability to correctly identify tumor boundaries and delineation consistency. In this study, we evaluated a novel Multisource Adaptive MRI Fusion (MAMF) method in HCC patients for tumor delineation. Methods: Ten patients with HCC were included in this study retrospectively. Contrast-enhanced T1-weighted MRI at portal-venous phase (T1WPP), contrast-enhanced T1-weighted MRI at 19-min delayed phase (T1WDP), T2-weighted (T2W), and diffusion-weighted MRI (DWI) were acquired on a 3T MRI scanner and imported to in-house-developed MAMF software to generate synthetic MR fusion images. The original multi-contrast MR image sets were registered to planning CT by deformable image registration (DIR) using MIM. Four observers independently delineated gross tumor volumes (GTVs) on the planning CT, four original MR image sets, and the fused MRI for all patients. Tumor contrast-to-noise ratio (CNR) and Dice similarity coefficient (DSC) of the GTVs between each observer and a reference observer were measured on the six image sets. Inter-observer and inter-patient mean, SD, and coefficient of variation (CV) of the DSC were evaluated. Results: Fused MRI showed the highest tumor CNR compared to planning CT and original MR sets in the ten patients. The mean ± SD tumor CNR was 0.72 ± 0.73, 3.66 ± 2.96, 4.13 ± 3.98, 4.10 ± 3.17, 5.25 ± 2.44, and 9.82 ± 4.19 for CT, T1WPP, T2W, DWI, T1WDP, and fused MRI, respectively. Fused MRI has the minimum inter-observer and inter-patient variations as compared to original MR sets and planning CT sets. GTV delineation inter-observer mean DSC across the ten patients was 0.81 ± 0.09, 0.85 ± 0.08, 0.88 ± 0.04, 0.89 ± 0.08, 0.90 ± 0.04, and 0.95 ± 0.02 for planning CT, T1WPP, T2W, DWI, T1WDP, and fused MRI, respectively. The patient mean inter-observer CV of DSC was 3.3%, 3.2%, 1.7%, 2.6%, 1.5%, and 0.9% for planning CT, T1WPP, T2W, DWI, T1WDP, and fused MRI, respectively. Conclusion: The results demonstrated that the fused MRI generated using the MAMF method can enhance tumor CNR and improve inter-observer consistency of GTV delineation in HCC as compared to planning CT and four commonly used MR image sets (T1WPP, T1WDP, T2W, and DWI). The MAMF method holds great promise in MRI applications in HCC radiotherapy treatment planning.

A Multi-Center Study of CT-Based Neck Nodal Radiomics for Predicting an Adaptive Radiotherapy Trigger of Ill-Fitted Thermoplastic Masks in Patients with Nasopharyngeal Carcinoma.

Significant lymph node shrinkage is common in patients with nasopharyngeal carcinoma (NPC) throughout radiotherapy (RT) treatment, causing ill-fitted thermoplastic masks (IfTMs). To deal with this, an ad hoc adaptive radiotherapy (ART) may be required to ensure accurate and safe radiation delivery and to maintain treatment efficacy. Presently, the entire procedure for evaluating an eligible ART candidate is time-consuming, resource-demanding, and highly inefficient. In the artificial intelligence paradigm, the pre-treatment identification of NPC patients at risk for IfTMs has become greatly demanding for achieving efficient ART eligibility screening, while no relevant studies have been reported. Hence, we aimed to investigate the capability of computed tomography (CT)-based neck nodal radiomics for predicting IfTM-triggered ART events in NPC patients via a multi-center setting. Contrast-enhanced CT and the clinical data of 124 and 58 NPC patients from Queen Elizabeth Hospital (QEH) and Queen Mary Hospital (QMH), respectively, were retrospectively analyzed. Radiomic (R), clinical (C), and combined (RC) models were developed using the ridge algorithm in the QEH cohort and evaluated in the QMH cohort using the median area under the receiver operating characteristics curve (AUC). Delong's test was employed for model comparison. Model performance was further assessed on 1000 replicates in both cohorts separately via bootstrapping. The R model yielded the highest "corrected" AUC of 0.784 (BCa 95%CI: 0.673-0.859) and 0.723 (BCa 95%CI: 0.534-0.859) in the QEH and QMH cohort following bootstrapping, respectively. Delong's test indicated that the R model performed significantly better than the C model in the QMH cohort (p < 0.0001), while demonstrating no significant difference compared to the RC model (p = 0.5773). To conclude, CT-based neck nodal radiomics was capable of predicting IfTM-triggered ART events in NPC patients in this multi-center study, outperforming the traditional clinical model. The findings of this study provide valuable insights for future study into developing an effective screening strategy for ART eligibility in NPC patients in the long run, ultimately alleviating the workload of clinical practitioners, streamlining ART procedural efficiency in clinics, and achieving personalized RT for NPC patients in the future.

Virtual Contrast-Enhanced Magnetic Resonance Images Synthesis for Patients With Nasopharyngeal Carcinoma Using Multimodality-Guided Synergistic Neural Network.

PURPOSE: To investigate a novel deep-learning network that synthesizes virtual contrast-enhanced T1-weighted (vceT1w) magnetic resonance images (MRI) from multimodality contrast-free MRI for patients with nasopharyngeal carcinoma (NPC). METHODS AND MATERIALS: This article presents a retrospective analysis of multiparametric MRI, with and without contrast enhancement by gadolinium-based contrast agents (GBCAs), obtained from 64 biopsy-proven cases of NPC treated at Hong Kong Queen Elizabeth Hospital. A multimodality-guided synergistic neural network (MMgSN-Net) was developed to leverage complementary information between contrast-free T1-weighted and T2-weighted MRI for vceT1w MRI synthesis. Thirty-five patients were randomly selected for model training, whereas 29 patients were selected for model testing. The synthetic images generated from MMgSN-Net were quantitatively evaluated against real GBCA-enhanced T1-weighted MRI using a series of statistical evaluating metrics, which include mean absolute error (MAE), mean squared error (MSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). Qualitative visual assessment between the real and synthetic MRI was also performed. Effectiveness of our MMgSN-Net was compared with 3 state-of-the-art deep-learning networks, including U-Net, CycleGAN, and Hi-Net, both quantitatively and qualitatively. Furthermore, a Turing test was performed by 7 board-certified radiation oncologists from 4 hospitals for assessing authenticity of the synthesized vceT1w MRI against the real GBCA-enhanced T1-weighted MRI. RESULTS: Results from the quantitative evaluations demonstrated that our MMgSN-Net outperformed U-Net, CycleGAN and Hi-Net, yielding the top-ranked scores in averaged MAE (44.50 ± 13.01), MSE (9193.22 ± 5405.00), SSIM (0.887 ± 0.042), and PSNR (33.17 ± 2.14). Furthermore, the mean accuracy of the 7 readers in the Turing tests was determined to be 49.43%, equivalent to random guessing (ie, 50%) in distinguishing between real GBCA-enhanced T1-weighted and synthetic vceT1w MRI. Qualitative evaluation indicated that MMgSN-Net gave the best approximation to the ground-truth images, particularly in visualization of tumor-to-muscle interface and the intratumor texture information. CONCLUSIONS: Our MMgSN-Net was capable of synthesizing highly realistic vceT1w MRI that outperformed the 3 comparable state-of-the-art networks.