A deep learning model for translating CT to ventilation imaging: analysis of accuracy and impact on functional avoidance radiotherapy planning.

PURPOSE: Radiotherapy planning incorporating functional lung images has the potential to reduce pulmonary toxicity. Free-breathing 4DCT-derived ventilation image (CTVI) may help quantify lung function. This study introduces a novel deep-learning model directly translating planning CT images into CTVI. We investigated the accuracy of generated images and the impact on functional avoidance planning. MATERIALS AND METHODS: Paired planning CT and 4DCT scans from 48 patients with NSCLC were randomized to training (n = 41) and testing (n = 7) data sets. The ventilation maps were generated from 4DCT using a Jacobian-based algorithm to provide ground truth labels (CTVI4DCT). A 3D U-Net-based model was trained to map CT to synthetic CTVI (CTVISyn) and validated using fivefold cross-validation. The highest-performing model was applied to the testing set. Spearman's correlation (rs) and Dice similarity coefficient (DSC) determined voxel-wise and functional-wise concordance between CTVI4DCT and CTVISyn. Three plans were designed per patient in the testing set: one clinical plan without CTVI and two functional avoidance plans combined with CTVI4DCT or CTVISyn, aimed at sparing high-functional lungs defined as the top 50% of the percentile ventilation ranges. Dose-volume (DVH) parameters regarding the planning target volume (PTV) and organs at risk (OARs) were recorded. Radiation pneumonitis (RP) risk was estimated using a dose-function (DFH)-based normal tissue complication probability (NTCP) model. RESULTS: CTVISyn showed a mean rs value of 0.65 ± 0.04 compared to CTVI4DCT. Mean DSC values over the top 50% and 60% of ventilation ranges were 0.41 ± 0.07 and 0.52 ± 0.10, respectively. In the test set (n = 7), all patients' RP-risk benefited from CTVI4DCT-guided plans (Riskmean_4DCT_vs_Clinical: 29.24% vs. 49.12%, P = 0.016), and six patients benefited from CTVISyn-guided plans (Riskmean_Syn_vs_Clinical: 31.13% vs. 49.12%, P = 0.022). There were no significant differences in DVH and DFH metrics between CTVISyn and CTVI4DCT-guided plan (P > 0.05). CONCLUSION: Using deep-learning techniques, CTVISyn generated from planning CT exhibited a moderate-to-high correlation with CTVI4DCT. The CTVISyn-guided plans were comparable to the CTVI4DCT-guided plans, effectively reducing pulmonary toxicity in patients while maintaining acceptable plan quality. Further prospective trials are needed to validate these findings.

Deep learning-based combination of [18F]-FDG PET and CT images for producing pulmonary perfusion image.

BACKGROUND: The main application of [18F] FDG-PET (18 FDG-PET) and CT images in oncology is tumor identification and quantification. Combining PET and CT images to mine pulmonary perfusion information for functional lung avoidance radiation therapy (FLART) is desirable but remains challenging. PURPOSE: To develop a deep-learning-based (DL) method to combine 18 FDG-PET and CT images for producing pulmonary perfusion images (PPI). METHODS: Pulmonary technetium-99 m-labeled macroaggregated albumin SPECT (PPISPECT ), 18 FDG-PET, and CT images obtained from 53 patients were enrolled. CT and PPISPECT images were rigidly registered, and registration displacement was subsequently used to align 18 FDG-PET and PPISPECT images. The left/right lung was separated and rigidly registered again to improve the registration accuracy. A DL model based on 3D Unet architecture was constructed to directly combine multi-modality 18 FDG-PET and CT images for producing PPI (PPIDLM ). 3D Unet architecture was used as the basic architecture, and the input was expanded from a single-channel to a dual-channel to combine multi-modality images. For comparative evaluation, 18 FDG-PET images were also used alone to generate PPIDLPET . Sixty-seven samples were randomly selected for training and cross-validation, and 36 were used for testing. The Spearman correlation coefficient (rs ) and multi-scale structural similarity index measure (MS-SSIM) between PPIDLM /PPIDLPET and PPISPECT were computed to assess the statistical and perceptual image similarities. The Dice similarity coefficient (DSC) was calculated to determine the similarity between high-/low- functional lung (HFL/LFL) volumes. RESULTS: The voxel-wise rs and MS-SSIM of PPIDLM /PPIDLPET were 0.78 ± 0.04/0.57 ± 0.03, 0.93 ± 0.01/0.89 ± 0.01 for cross-validation and 0.78 ± 0.11/0.55 ± 0.18, 0.93 ± 0.03/0.90 ± 0.04 for testing. PPIDLM /PPIDLPET achieved averaged DSC values of 0.78 ± 0.03/0.64 ± 0.02 for HFL and 0.83 ± 0.01/0.72 ± 0.03 for LFL in the training dataset and 0.77 ± 0.11/0.64 ± 0.12, 0.82 ± 0.05/0.72 ± 0.06 in the testing dataset. PPIDLM yielded a stronger correlation and higher MS-SSIM with PPISPECT than PPIDLPET (p < 0.001). CONCLUSIONS: The DL-based method integrates lung metabolic and anatomy information for producing PPI and significantly improved the accuracy over methods based on metabolic information alone. The generated PPIDLM can be applied for pulmonary perfusion volume segmentation, which is potentially beneficial for FLART treatment plan optimization.

Development and verification of radiomics framework for computed tomography image segmentation.

BACKGROUND: Radiomics has been considered an imaging marker for capturing quantitative image information (QII). The introduction of radiomics to image segmentation is desirable but challenging. PURPOSE: This study aims to develop and validate a radiomics-based framework for image segmentation (RFIS). METHODS: RFIS is designed using features extracted from volume (svfeatures) created by sliding window (swvolume). The 53 svfeatures are extracted from 11 phantom series. Outliers in the svfeature datasets are detected by isolation forest (iForest) and specified as the mean value. The percentage coefficient of variation (%COV) is calculated to evaluate the reproducibility of svfeatures. RFIS is constructed and applied to the gross target volume (GTV) segmentation from the peritumoral region (GTV with a 10 mm margin) to assess its feasibility. The 127 lung cancer images are enrolled. The test-retest method, correlation matrix, and Mann-Whitney U test (p < 0.05) are used to select non-redundant svfeatures of statistical significance from the reproducible svfeatures. The synthetic minority over-sampling technique is utilized to balance the minority group in the training sets. The support vector machine is employed for RFIS construction, which is tuned in the training set using 10-fold stratified cross-validation and then evaluated in the test sets. The swvolumes with the consistent classification results are grouped and merged. Mode filtering is performed to remove very small subvolumes and create relatively large regions of completely uniform character. In addition, RFIS performance is evaluated by the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and Dice similarity coefficient (DSC). RESULTS: 30249 phantom and 145008 patient image swvolumes were analyzed. Forty-nine (92.45% of 53) svfeatures represented excellent reproducibility(%COV<15). Forty-five features (91.84% of 49) included five categories that passed test-retest analysis. Thirteen svfeatures (28.89% of 45) svfeatures were selected for RFIS construction. RFIS showed an average (95% confidence interval) sensitivity of 0.848 (95% CI:0.844-0.883), a specificity of 0.821 (95% CI: 0.818-0.825), an accuracy of 83.48% (95% CI: 83.27%-83.70%), and an AUC of 0.906 (95% CI: 0.904-0.908) with cross-validation. The sensitivity, specificity, accuracy, and AUC were equal to 0.762 (95% CI: 0.754-0.770), 0.840 (95% CI: 0.837-0.844), 82.29% (95% CI: 81.90%-82.60%), and 0.877 (95% CI: 0.873-0.881) in the test set, respectively. GTV was segmented by grouping and merging swvolume with identical classification results. The mean DSC after mode filtering was 0.707 ± 0.093 in the training sets and 0.688 ± 0.072 in the test sets. CONCLUSION: Reproducible svfeatures can capture the differences in QII among swvolumes. RFIS can be applied to swvolume classification, which achieves image segmentation by grouping and merging the swvolume with similar QII.

Predicting HER2 Status in Breast Cancer on Ultrasound Images Using Deep Learning Method.

PURPOSE: The expression of human epidermal growth factor receptor 2 (HER2) in breast cancer is critical in the treatment with targeted therapy. A 3-block-DenseNet-based deep learning model was developed to predict the expression of HER2 in breast cancer by ultrasound images. METHODS: The data from 144 breast cancer patients with preoperative ultrasound images and clinical information were retrospectively collected from the Shandong Province Tumor Hospital. An end-to-end 3-block-DenseNet deep learning classifier was built to predict the expression of human epidermal growth factor receptor 2 by ultrasound images. The patients were randomly divided into a training (n = 108) and a validation set (n = 36). RESULTS: Our proposed deep learning model achieved an encouraging predictive performance in the training set (accuracy = 85.79%, AUC = 0.87) and the validation set (accuracy = 80.56%, AUC = 0.84). The effectiveness of our model significantly exceeded the clinical model and the radiomics model. The score of the proposed model showed significant differences between HER2-positive and -negative expression (p < 0.001). CONCLUSIONS: These results demonstrate that ultrasound images are predictive of HER2 expression through a deep learning classifier. Our method provides a non-invasive, simple, and feasible method for the prediction of HER2 expression without the manual delineation of the regions of interest (ROI). The performance of our deep learning model significantly exceeded the traditional texture analysis based on the radiomics model.

The Capabilities and Characteristics of Helical Tomotherapy and Co-Planar Dual Arcs Volumetric-Modulated arc Therapy Associated with Hippocampal Sparing During Prophylactic Cranial Irradiation.

Objective: To investigate the features of helical tomotherapy and co-planar dual Arcs volumetric-modulated arc therapy during prophylactic cranial irradiation associated with bilateral hippocampal tissue sparing. Materials and methods: Helical tomotherapy and co-planar dual arcs volumetric-modulated arc therapy treatment plans were generated with a dose of 30 Gy/10 fractions in 16 patients treated with prophylactic cranial irradiation. The dose to the bilateral hippocampal tissues, organs at risk, and planning target volume were determined when the average dose of bilateral hippocampal tissues was reduced by approximately 4 Gy as an observation point. Changes in dosimetry when sparing the bilateral hippocampal tissues were determined for both modalities. Results: When bilateral hippocampal tissues were restricted to 8 Gy, D40%mean-bilateral hippocampal tissues = 7.64 ± 0.41 Gy in helical tomotherapy, while D40%mean-bilateral hippocampal tissues = 10.96 ± 0.38 Gy in co-planar dual arcs volumetric-modulated arc therapy volumetric-modulated arc therapy. Helical tomotherapy was associated with significantly lower doses to organs at risk, including Dmean-bilateral hippocampal tissues (P = .03), D98%-bilateral hippocampal tissues (P = .01), D2%-bilateral hippocampal tissues (P = .01), Dmean-inner ear (P = .02), Dmean-parotid glands (P = .02), Dmax-lens (P = .02), and Dmax-brainstem (P = .02), but not Dmax-optic nerves (P = .87). Helical tomotherapy provided better target coverage, with lower average D2%-PTV (P = .02), higher average D98%-PTV (P = .02), and better conformal index (0.87 vs 0.84, P = .02) and homogeneity index (0.15 vs 0.21, P = .05). With smaller bilateral hippocampal tissues doses, the planning target volume dose changed across 3 dosimetry regions for both modalities; the plateau region (>20.0 Gy for helical tomotherapy versus >16.0 Gy for co-planar dual arcs volumetric-modulated arc therapy), gradient region (20.0-12.0 Gy vs 16.0-11.0 Gy), and falling region (<12.0 Gy vs <11.0 Gy). The average delivery duration of helical tomotherapy was almost 7.7 times longer than that of co-planar dual arcs volumetric-modulated arc therapy. Conclusions: Helical tomotherapy was better at sparing the bilateral hippocampal tissues and organs at risk and had better target coverage but a significantly longer treatment duration than co-planar dual arcs volumetric-modulated arc therapy. Further dose decreases in the bilateral hippocampal tissues would yield worse target dose coverage.

Development and Validation of a Radiomics Nomogram Model for Predicting Postoperative Recurrence in Patients With Esophageal Squamous Cell Cancer Who Achieved pCR After Neoadjuvant Chemoradiotherapy Followed by Surgery.

Background and purpose: Although patients with esophageal squamous cell carcinoma (ESCC) can achieve a pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) followed by surgery, one-third of these patients with a pCR may still experience recurrence. The aim of this study is to develop and validate a predictive model to estimate recurrence-free survival (RFS) in those patients who achieved pCR. Materials and methods: Two hundred six patients with ESCC were enrolled and divided into a training cohort (n = 146) and a validation cohort (n = 60). Radiomic features were extracted from contrast-enhanced computed tomography (CT) images of each patient. Feature reduction was then implemented in two steps, including a multiple segmentation test and least absolute shrinkage and selection operator (LASSO) Cox proportional hazards regression method. A radiomics signature was subsequently constructed and evaluated. For better prediction performance, a clinical nomogram based on clinical risk factors and a nomogram incorporating the radiomics signature and clinical risk factors was built. Finally, the prediction models were further validated by calibration and the clinical usefulness was examined in the validation cohort to determine the optimal prediction model. Results: The radiomics signature was constructed using eight radiomic features and displayed a significant correlation with RFS. The nomogram incorporating the radiomics signature with clinical risk factors achieved optimal performance compared with the radiomics signature (P < 0.001) and clinical nomogram (P < 0.001) in both the training cohort [C-index (95% confidence interval [CI]), 0.746 (0.680-0.812) vs. 0.685 (0.620-0.750) vs. 0.614 (0.538-0.690), respectively] and validation cohort [C-index (95% CI), 0.724 (0.696-0.752) vs. 0.671 (0.624-0.718) vs. 0.629 (0.597-0.661), respectively]. The calibration curve and decision curve analysis revealed that the radiomics nomogram outperformed the other two models. Conclusions: A radiomics nomogram model incorporating radiomics features and clinical factors has been developed and has the improved ability to predict the postoperative recurrence risk in patients with ESCC who achieved pCR after nCRT followed by surgery.

Prediction of Immunohistochemistry of Suspected Thyroid Nodules by Use of Machine Learning-Based Radiomics.

OBJECTIVE. The purpose of this study was to develop and validate a radiomics model for evaluating immunohistochemical characteristics in patients with suspected thyroid nodules. MATERIALS AND METHODS. A total of 103 patients (training cohort-to-validation cohort ratio, ≈ 3:1) with suspected thyroid nodules who had undergone thyroidectomy and immunohistochemical analysis were enrolled. The immunohistochemical markers were cytokeratin 19, galectin 3, thyroperoxidase, and high-molecular-weight cytokeratin. All patients underwent CT before surgery, and a 3D slicer was used to analyze images of the surgical specimen. Test-retest and Spearman correlation coefficient (ρ) were used to select reproducible and nonredundant features. The Kruskal-Wallis test (p < 0.05) was used for feature selection, and a feature-based model was built by support vector machine methods. The performance of the radiomic models was assessed with respect to accuracy, sensitivity, specificity, corresponding AUC, and independent validation. RESULTS. Eighty-six reproducible and nonredundant features selected from the 828 features were used to build the model. The best performance of the cytokeratin 19 model yielded accuracy of 84.4% in the training cohort and 80.0% in the validation cohort. The thyroperoxidase and galectin 3 predictive models yielded accuracies of 81.4% and 82.5% in the training cohort and 84.2% and 85.0% in the validation cohort. The performance of the high-molecular-weight cytokeratin predictive model was not good (accuracy, 65.7%) and could not be validated. CONCLUSION. A radiomics model with excellent performance was developed for individualized noninvasive prediction of the presence of cytokeratin 19, galectin 3, and thyroperoxidase based on CT images. This model may be used to identify benign and malignant thyroid nodules.

The Feasibility Study of Megavoltage Computed Tomographic (MVCT) Image for Texture Feature Analysis.

Purpose: To determine whether radiomics texture features can be reproducibly obtained from megavoltage computed tomographic (MVCT) images acquired by Helical TomoTherapy (HT) with different imaging conditions. Methods: For each of the 195 textures enrolled, the mean intrapatient difference, which is considered to be the benchmark for reproducibility, was calculated from the MVCT images of 22 patients with early-stage non-small-cell lung cancer. Test-retest MVCT images of an in-house designed phantom were acquired to determine the concordance correlation coefficient (CCC) for these 195 texture features. Features with high reproducibility (CCC > 0.9) in the phantom test-retest set were investigated for sensitivities to different imaging protocols, scatter levels, and motion frequencies using a wood phantom and in-vitro animal tissues. Results: Of the 195 features, 165 (85%) features had CCC > 0.9. For the wood phantom, 124 features were reproducible in two kinds of scatter materials, and further investigations were performed on these features. For animal tissues, 108 features passed the criteria for reproducibility when one layer of scatter was covered, while 106 and 108 features of in-vitro liver and bone passed with two layers of scatter, respectively. Considering the effect of differing acquisition pitch (AcP), 97 features extracted from wood passed, while 103 and 59 features extracted from in-vitro liver and bone passed, respectively. Different reconstruction intervals (RI) had a small effect on the stability of the feature value. When AcP and RI were held consistent without motion, all 124 features calculated from wood passed, and a majority (122 of 124) of the features passed when imaging with a "fine" AcP with different RIs. However, only 55 and 40 features passed with motion frequencies of 20 and 25 beats per minute, respectively. Conclusion: Motion frequency has a significant impact on MVCT texture features, and features from MVCT were more reproducibility in different scatter conditions than those from CBCT. Considering the effects of AcP and RI, the scanning protocols should be kept consistent when MVCT images are used for feature analysis. Some radiomics features from HT MVCT images are reproducible and could be used for creating clinical prediction models in the future.

Computational delineation and quantitative heterogeneity analysis of lung tumor on 18F-FDG PET for radiation dose-escalation.

Quantitative measurement and analysis of tumor metabolic activities could provide a more optimal solution to personalized accurate dose painting. We collected PET images of 58 lung cancer patients, in which the tumor exhibits heterogeneous FDG uptake. We design an automated delineation and quantitative heterogeneity measurement of the lung tumor for dose-escalation. For tumor delineation, our algorithm firstly separates the tumor from its adjacent high-uptake tissues using 3D projection masks; then the tumor boundary is delineated with our stopping criterion of joint gradient and intensity affinities. For dose-escalation, tumor sub-volumes with low, moderate and high metabolic activities are extracted and measured. Based on our quantitative heterogeneity measurement, a sub-volume oriented dose-escalation plan is implemented in intensity modulated radiation therapy (IMRT) planning system. With respect to manual tumor delineations by two radiation oncologists, the paired t-test demonstrated our model outperformed the other computational methods in comparison (p < 0.05) and reduced the variability between inter-observers. Compared to standard uniform dose prescription, the dosimetry results demonstrated that the dose-escalation plan statistically boosted the dose delivered to high metabolic tumor sub-volumes (p < 0.05). Meanwhile, the doses received by organs-at-risk (OAR) including the heart, ipsilateral lung and contralateral lung were not statistically different (p > 0.05).

Effects of megavoltage computed tomographic scan methodology on setup verification and adaptive dose calculation in helical TomoTherapy.

BACKGROUND: To evaluate the effect of pretreatment megavoltage computed tomographic (MVCT) scan methodology on setup verification and adaptive dose calculation in helical TomoTherapy. METHODS: Both anthropomorphic heterogeneous chest and pelvic phantoms were planned with virtual targets by TomoTherapy Physicist Station and were scanned with TomoTherapy megavoltage image-guided radiotherapy (IGRT) system consisted of six groups of options: three different acquisition pitches (APs) of 'fine', 'normal' and 'coarse' were implemented by multiplying 2 different corresponding reconstruction intervals (RIs). In order to mimic patient setup variations, each phantom was shifted 5 mm away manually in three orthogonal directions respectively. The effect of MVCT scan options was analyzed in image quality (CT number and noise), adaptive dose calculation deviations and positional correction variations. RESULTS: MVCT scanning time with pitch of 'fine' was approximately twice of 'normal' and 3 times more than 'coarse' setting, all which will not be affected by different RIs. MVCT with different APs delivered almost identical CT numbers and image noise inside 7 selected regions with various densities. DVH curves from adaptive dose calculation with serial MVCT images acquired by varied pitches overlapped together, where as there are no significant difference in all p values of intercept & slope of emulational spinal cord (p = 0.761 & 0.277), heart (p = 0.984 & 0.978), lungs (p = 0.992 & 0.980), soft tissue (p = 0.319 & 0.951) and bony structures (p = 0.960 & 0.929) between the most elaborated and the roughest serials of MVCT. Furthermore, gamma index analysis shown that, compared to the dose distribution calculated on MVCT of 'fine', only 0.2% or 1.1% of the points analyzed on MVCT of 'normal' or 'coarse' do not meet the defined gamma criterion. On chest phantom, all registration errors larger than 1 mm appeared at superior-inferior axis, which cannot be avoided with the smallest AP and RI. On pelvic phantom, craniocaudal errors are much smaller than chest, however, AP of 'coarse' presents larger registration errors which can be reduced from 2.90 mm to 0.22 mm by registration technique of 'full image'. CONCLUSIONS: AP of 'coarse' with RI of 6 mm is recommended in adaptive radiotherapy (ART) planning to provide craniocaudal longer and faster MVCT scan, while registration technique of 'full image' should be used to avoid large residual error. Considering the trade-off between IGRT and ART, AP of 'normal' with RI of 2 mm was highly recommended in daily practice.

A method for estimating spikelet number per panicle: Integrating image analysis and a 5-point calibration model.

Spikelet number per panicle (SNPP) is one of the most important yield components used to estimate rice yields. The use of high-throughput quantitative image analysis methods for understanding the diversity of the panicle has increased rapidly. However, it is difficult to simultaneously extract panicle branch and spikelet/grain information from images at the same resolution due to the different scales of these traits. To use a lower resolution and meet the accuracy requirement, we proposed an interdisciplinary method that integrated image analysis and a 5-point calibration model to rapidly estimate SNPP. First, a linear relationship model between the total length of the primary branch (TLPB) and the SNPP was established based on the physiological characteristics of the panicle. Second, the TLPB and area (the primary branch region) traits were rapidly extracted by developing image analysis algorithm. Finally, a 5-point calibration method was adopted to improve the universality of the model. The number of panicle samples that the error of the SNPP estimates was less than 10% was greater than 90% by the proposed method. The estimation accuracy was consistent with the accuracy determined using manual measurements. The proposed method uses available concepts and techniques for automated estimations of rice yield information.