Deep Learning Algorithm for Reducing CT Slice Thickness: Effect on Reproducibility of Radiomic Features in Lung Cancer

Background: The reproducibility of radiomic features is a critical challenge facing radiomic models of tumor prediction or prognosis. The aim of this study is to evaluate the reproducibility of radiomic features with the GrowCut and GraphCut semi-automatic tumor segmentation methods in hepatocellular carcinoma (HCC) CT images. Methods: Computed tomography (CT) data sets (arterial enhanced phase) of 15 patients with HCC were randomly selected in this study. To acquire the gross tumor volume (GTV), semi-automatic segmentation with the GrowCut and GraphCut methods was implemented in 3D Slicer software by two independent observers. Meanwhile, manual delineation of the GTV was implemented by five abdomen radiation oncologists in this study. We divided the sample into three groups: the GrowCut group, the GraphCut group and manual group. Radiomic features (including tumor intensity histogram-based features, textural features and shape-based features) were extracted using the Imaging Biomarker Explorer (IBEX) software. The intraclass correlation coefficient (ICC) was applied to assess the reproducibility of all radiomic features. Results: The radiomic features in the GrowCut group (ICC =0.87±0.19) showed higher reproducibility compared with the radiomic features in the GraphCut group (ICC =0.82±0.24, P Conclusions: Our study reveals that variations exist in the reproducibility of quantitative imaging features extracted from tumor regions segmented using different methods. The 3D Slicer can serve as a better alternative to the manual delineation, and care must be taken when selecting segmentation tools to draw the tumor region.

PurposeTo improve reproducibility and predictive performance of PET radiomic features in multicentric studies by cycle-consistent generative adversarial network (GAN) harmonization approaches.MethodsGAN-harmonization was developed to harmonize whole-body PET scans to perform image style and texture translation between different centers and scanners. GAN-harmonization was evaluated by application to two retrospectively collected open datasets and different tasks. First, GAN-harmonization was performed on a dual-center lung cancer cohort (127 female, 138 male) where the reproducibility of radiomic features in healthy liver tissue was evaluated. Second, GAN-harmonization was applied to a head and neck cancer cohort (43 female, 154 male) acquired from three centers. Here, the clinical impact of GAN-harmonization was analyzed by predicting the development of distant metastases using a logistic regression model incorporating first-order statistics and texture features from baseline 18F-FDG PET before and after harmonization.ResultsImage quality remained high (structural similarity: left kidney ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.800, right kidney ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.806, liver ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.780, lung ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.838, spleen ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.793, whole-body ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 0.832) after image harmonization across all utilized datasets. Using GAN-harmonization, inter-site reproducibility of radiomic features in healthy liver tissue increased at least by ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 5 ± 14% (first-order), ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 16 ± 7% (GLCM), ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 19 ± 5% (GLRLM), ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 16 ± 8% (GLSZM), ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 17 ± 6% (GLDM), and ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge$$\\end{document} 23 ± 14% (NGTDM). In the head and neck cancer cohort, the outcome prediction improved from AUC 0.68 (95% CI 0.66–0.71) to AUC 0.73 (0.71–0.75) by application of GAN-harmonization.ConclusionsGANs are capable of performing image harmonization and increase reproducibility and predictive performance of radiomic features derived from different centers and scanners.

The impact of respiratory motion and CT pitch on the robustness of radiomics feature extraction in 4DCT lung imaging

Imaging phantoms were scanned twice on 3 computed tomography scanners from 2 different manufactures with varying tube voltages and currents. Phantoms were segmented, and features were extracted using PyRadiomics and a pretrained CNN. After standardization the concordance correlation coefficient (CCC), mean feature variance, feature range, and the coefficient of variant were calculated to assess feature robustness. In addition, the cosine similarity was calculated for the vectorized activation maps for an exemplary phantom. For the in vivo comparison, the radiomics and CNN features of 30 patients with hepatocellular carcinoma (HCC) and 30 patients with hepatic colon carcinoma metastasis were compared. In total, 851 radiomics features and 256 CNN features were extracted for each phantom. For all phantoms, the global CCC of the CNN features was above 98%, whereas the highest CCC for the radiomics features was 36%. The mean feature variance and feature range was significantly lower for the CNN features. Using a coefficient of variant ≤0.2 as a threshold to define robust features and averaging across all phantoms 346 of 851 (41%) radiomics features and 196 of 256 (77%) CNN features were found to be robust. The cosine similarity was greater than 0.98 for all scanner and parameter variations. In the retrospective analysis, 122 of the 256 CNN (49%) features showed significant differences between HCC and hepatic colon metastasis. Convolutional neural network features were more stable compared with radiomics features against technical variations. Moreover, the possibility of tumor entity differentiation based on CNN features was shown. Combined with visualization methods, CNN features are expected to increase reproducibility of quantitative image representations. Further studies are warranted to investigate the impact of feature stability on radiological image-based prediction of clinical outcomes.

Introduction: The purpose of this study is to investigate the effects of automatic segmentation algorithms on the performance of ultrasound (US) radiomics models in predicting the status of lymph node metastasis (LNM) for patients with early stage cervical cancer preoperatively. Methods: US images of 148 cervical cancer patients were collected and manually contoured by two senior radiologists. The four deep learning-based automatic segmentation models, namely U-net, context encoder network (CE-net), Resnet, and attention U-net were constructed to segment the tumor volumes automatically. Radiomics features were extracted and selected from manual and automatically segmented regions of interest (ROIs) to predict the LNM of these cervical cancer patients preoperatively. The reliability and reproducibility of radiomics features and the performances of prediction models were evaluated. Results: A total of 449 radiomics features were extracted from manual and automatic segmented ROIs with Pyradiomics. Features with an intraclass coefficient (ICC) > 0.9 were all 257 (57.2%) from manual and automatic segmented contours. The area under the curve (AUCs) of validation models with radiomics features extracted from manual, attention U-net, CE-net, Resnet, and U-net were 0.692, 0.755, 0.696, 0.689, and 0.710, respectively. Attention U-net showed best performance in the LNM prediction model with a lowest discrepancy between training and validation. The AUCs of models with automatic segmentation features from attention U-net, CE-net, Resnet, and U-net were 9.11%, 0.58%, –0.44%, and 2.61% higher than AUC of model with manual contoured features, respectively. Conclusion: The reliability and reproducibility of radiomics features, as well as the performance of radiomics models, were affected by manual segmentation and automatic segmentations.

Dose-volume histogram (DVH) has become an important tool for evaluation of radiation outcome as reflected from many clinical protocols. While dosimetric accuracy in treatment planning system (TPS) is well quantified, the variability in volume estimation is uncertain due to reconstruction algorithm that is investigated in this study. In addition, the impact of dose distribution and tumor control probability (TCP) were also investigated with CT slice thickness for IMRT planning. A water phantom containing various objects with accurately known volume ranging from 1 to 100 cm(3) was scanned with 1, 2, 3, 5, and 10 mm slice thickness. The CT data sets were sent to Eclipse TPS for contour delineation and volume estimation. The data were compared with known volume for the estimation of error in the volume of each structure. IMRT Plans were generated on phantom containing four objects with different slice thickness (1-5 mm) to calculate TCP. ICRU-83-recommended dose points such as D 2%, D 50%, D 98%, as well as homogeneity and conformity index were also calculated. The variability of volumes with CT slice thickness was significant especially for small volume structures. A maximum error of 92% was noticed for 1 cm(3) volume of object with 10 mm slice thickness, whereas it was ~19% for 1 mm slice thickness. For 2 and 3 cm(3) objects, the maximum error of 99% was noticed with 10 mm slice thickness and ~60% with 5 mm. The differences are smaller for larger volumes with a cutoff at about 20 cm(3). The calculated volume of the objects is a function of reconstruction algorithm and slice thickness. The PTV mean dose and TCP decreased with increasing slice thickness. Maximum variation of ~5% was noticed in mean dose and ~2% in TCP with change in slice thickness from 1 to 5 mm. The relative decrease in target volume receiving 95% of the prescribed dose is ~5% with change in slice thickness from 1 to 5 mm. The homogeneity index increases up to 163% and conformity index decreases by 4% between 1 and 5 mm slice thickness, producing highly inhomogeneous and least conformal treatment plan. Estimation of a volume is dependent on CT slice thickness and the contouring algorithm in a TPS. During commissioning of TPS and for all clinical protocols, evaluation of volume should be included to provide the limit of accuracy in DVH from TPS, especially for small objects. A smaller slice thickness provides superior dosimetry with improved TCP. Thus, the smallest possible slice thickness should be used for IMRT planning, especially when smaller structures are present.

Volumetric and whole-tumor attenuation assessment of tumor are of value in assessment of treatment. To assess the impact of slice thickness on semi-automatic analyses (volume, whole-tumor attenuation) for small colorectal hepatic metastases. Computed tomography (CT) data of patients with colorectal hepatic metastases at 1.5-, 3-, and 5-mm slice thickness were semi-automatically evaluated for volume and whole-tumor attenuation by two radiologists independently. Statistical analysis included paired samples t-test and concordance correlation coefficient (CCC) analysis according to the longest axial tumor diameter (10-20 mm, 20-30 mm, 30-40 mm). A total of 62 patients (32 men and 30 women) with 62 target tumors were included. The mean volume was significantly higher at 3- and 5-mm slice thicknesses in comparison with the reference (1.5 mm) for the target tumors between 10 mm and 20 mm (P = 0.0295, CCC = 0.9394 for 3 mm; P = 0.0029, CCC = 0.5129 for 5 mm, respectively) and at 5 mm slice thickness for the target tumors between 20 mm and 30 mm (P = 0.0071, CCC = 0.9102). For whole-tumor attenuation measurements, the significant difference was only seen at 5-mm slice thicknesses in comparison with the reference (1.5 mm) for the target tumors between 10 and 20 mm (P = 0.0015, CCC = 0.9389). Slice thickness of 1.5 mm might be suggested for semi-automated volumetric measurements, and slice thickness of no more than 3 mm for whole-tumor CT attenuation in hepatic metastasis between 10 mm and 20 mm.

Objective. The study aims to systematically characterize the effect of CT parameter variations on images and lung radiomic and deep features, and to evaluate the ability of different image harmonization methods to mitigate the observed variations.Approach. A retrospective in-house sinogram dataset of 100 low-dose chest CT scans was reconstructed by varying radiation dose (100%, 25%, 10%) and reconstruction kernels (smooth, medium, sharp). A set of image processing, convolutional neural network (CNNs), and generative adversarial network-based (GANs) methods were trained to harmonize all image conditions to a reference condition (100% dose, medium kernel). Harmonized scans were evaluated for image similarity using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS), and for the reproducibility of radiomic and deep features using concordance correlation coefficient (CCC).Main Results. CNNs consistently yielded higher image similarity metrics amongst others; for Sharp/10%, which exhibited the poorest visual similarity, PSNR increased from a mean ± CI of 17.763 ± 0.492 to 31.925 ± 0.571, SSIM from 0.219 ± 0.009 to 0.754 ± 0.017, and LPIPS decreased from 0.490 ± 0.005 to 0.275 ± 0.016. Texture-based radiomic features exhibited a greater degree of variability across conditions, i.e. a CCC of 0.500 ± 0.332, compared to intensity-based features (0.972 ± 0.045). GANs achieved the highest CCC (0.969 ± 0.009 for radiomic and 0.841 ± 0.070 for deep features) amongst others. CNNs are suitable if downstream applications necessitate visual interpretation of images, whereas GANs are better alternatives for generating reproducible quantitative image features needed for machine learning applications.Significance. Understanding the efficacy of harmonization in addressing multi-parameter variability is crucial for optimizing diagnostic accuracy and a critical step toward building generalizable models suitable for clinical use.

Computer-Aided-Diagnosis (CADx) systems assist radiologists with identifying and classifying potentially malignant pulmonary nodules on chest CT scans using morphology and texture-based (radiomic) features. However, radiomic features are sensitive to differences in acquisitions due to variations in dose levels and slice thickness. This study investigates the feasibility of generating a normalized scan from heterogeneous CT scans as input. We obtained projection data from 40 low-dose chest CT scans, simulating acquisitions at 10%, 25% and 50% dose and reconstructing the scans at 1.0mm and 2.0mm slice thickness. A 3D generative adversarial network (GAN) was used to simultaneously normalize reduced dose, thick slice (2.0mm) images to normal dose (100%), thinner slice (1.0mm) images. We evaluated the normalized image quality using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS). Our GAN improved perceptual similarity by 35%, compared to a baseline CNN method. Our analysis also shows that the GAN-based approach led to a significantly smaller error (p-value <; 0.05) in nine studied radiomic features. These results indicated that GANs could be used to normalize heterogeneous CT images and reduce the variability in radiomic feature values.

Independent and reproducible hippocampal radiomics biomarkers for multisite multiple sclerosis and neuromyelitis optica spectrum disorders

Radiomics features (RFs) studies have showed limitations in the reproducibility of RFs in different acquisition settings. To date, reproducibility studies using CT images mainly rely on phantoms, due to the harness of patient exposure to X-rays. The provided CadAIver dataset has the aims of evaluating how CT scanner parameters effect radiomics features on cadaveric donor. The dataset comprises 112 unique CT acquisitions of a cadaveric truck acquired on 3 different CT scanners varying KV, mA, field-of-view, and reconstruction kernel settings. Technical validation of the CadAIver dataset comprises a comprehensive univariate and multivariate GLM approach to assess stability of each RFs extracted from lumbar vertebrae. The complete dataset is publicly available to be applied for future research in the RFs field, and could foster the creation of a collaborative open CT image database to increase the sample size, the range of available scanners, and the available body districts.

CT Tumor Volume Measurement in Advanced Non-small-cell Lung Cancer: Performance Characteristics of an Emerging Clinical Tool

Purpose: In our experience, both recent versions (7.1 and 8.0) of the Variseed software have shown variability in seed detection capabilities—wrong identification, missing seeds. A phantom study was begun to improve our understanding of the effectiveness of seed identification in CT scans of patients following prostate implantation. Materials and Methods: A simple planar phantom was constructed for investigation of the effects of seed spacing, orientation, and CT slice thickness on the ability of commercial software to detect the positions and numbers of seeds accurately. Single seed spacings were varied from 4 mm to 11 mm in the longitudinal direction and from 1.9 to 8 mm in the axial plane; stranded seeds (10 mm spacing) were used to test the effect of plane orientations of 90, 75, and 60 deg (relative to the axial CT plane); CT slice thicknesses were varied from 1 to 5 mm; two versions of Varian's seed finder software (7.1 and 8.0) were compared for loose seed detection. V8.0 was used for the angular/slice thickness study. Non‐radioactive STM 1251 seeds were supplied by Bard Brachytherapy, Inc. Results: V 7.1 was slightly superior to V8.0 for loose seed identification in the longitudinal (100% vs 93%), but inferior (13% vs 98%) in the cross direction. No effect of slice thickness was observed. Increasing the angle between the seed plane and CT cross plane tended to increase both mean and maximum errors in seed spacing (up to 5 mm). Slice thicknesses of 2 or 3 mm seemed optimal. Conclusions: The seed phantom revealed differences between the two versions of the Variseed software for seed detection, a dependence on seed plane angle and confirmed the recommended CT slice thickness of 2–3 mm (TG137).

Introduction. In oncology, accurate classification of lung cancer mutations plays a key role in developing personalized treatment strategies. Lung cancer, distinguished by its heterogeneity, presents significant challenges in diagnosis and treatment, requiring innovative approaches for precise mutation classification.Aim. To introduce a new methodology combining deep learning and radiomic features extracted from computed tomography (CT) images for classification of lung cancer mutations.Materials and methods. The ResNet18 architecture was adapted to integrate radiomic features directly into the deep learning workflow. The use of a convolutional neural network enabled large volumes of data to be processed, surpassing the performance of conventional methods. The analysis involved identification of significant radiomic features, such as texture, shape, and tumor boundaries, which were automatically extracted and used to train the model. The technique was tested on an extensive dataset containing CT images of various lung cancer subtypes, including adenocarcinoma and squamous cell carcinoma.Results. The model demonstrated an overall mutation classification accuracy of 98.6 %, significantly exceeding the results achieved using conventional approaches. The high accuracy confirms the effectiveness of combining radiomic features with deep learning in identifying various genetic mutations in lung cancer. The results also indicate the high potential of the method in the development of non-invasive diagnostic tools and improving personalized treatment approaches.Conclusion. This work emphasizes the importance of integrating high-level abstractions of deep learning models with detailed analysis of radiomic data to enhance the predictive accuracy of non-invasive diagnostic tools, which could significantly improve diagnostic processes and contribute to the development of treatment strategies in oncology.

To predict the epidermal growth factor receptor (EGFR) mutation status in patients with lung adenocarcinoma using quantitative radiomic biomarkers and semantic features.We analyzed the computed tomography (CT) images and medical record data of 104 patients with lung adenocarcinoma who underwent surgical excision and EGFR mutation detection from 2016 to 2018 at our center. CT radiomic and semantic features that reflect the tumors’ heterogeneity and phenotype were extracted from preoperative non-enhanced CT scans. The least absolute shrinkage and selection operator method was applied to select the most distinguishable features. Three logistic regression models were built to predict the EGFR mutation status by combining the CT semantic with clinicopathological characteristics, using the radiomic features alone, and by combining the radiomic and clinicopathological features. Receiver operating characteristic (ROC) curve analysis was performed using five-fold cross-validation and the mean area under the curve (AUC) values were calculated and compared between the models to obtain the optimal model for predicting EGFR mutation. Furthermore, radiomic nomograms were constructed to demonstrate the performance of the model.In total, 1025 radiomic features were extracted and reduced to 13 features as the most important predictors to build the radiomic signature. The combined radiomic and clinicopathological features model was developed based on the radiomic signature, sex, smoking, vascular infiltration, and pathohistological type. The AUC was 0.90 ± 0.02 for the training, 0.88 ± 0.11 for the verification, and 0.894 for the test dataset. This model was superior to the other prediction models that used the combined CT semantic and clinicopathological features (AUC for the test dataset: 0.768) and radiomic features alone (AUC for the test dataset: 0.837).The prediction model built by radiomic biomarkers and clinicopathological features, including the radiomic signature, sex, smoking, vascular infiltration, and pathological type, outperformed the other two models and could effectively predict the EGFR mutation status in patients with peripheral lung adenocarcinoma. The radiomic nomogram of this model is expected to become an effective biomarker for patients with lung adenocarcinoma requiring adjuvant targeted treatment.