Deep Learning-Based Image Conversion Improves the Reproducibility of Computed Tomography Radiomics Features: A Phantom Study

Quantitative Liver Imaging in Children

ABSTRACT

In children and adults, quantitative imaging examinations determine the effectiveness of treatment for liver disease. However, pediatric liver disease differs in presentation from liver disease in adults. Children also needed to be followed for a longer period from onset and have less control of their bodies, showing more movement than adults during imaging examinations, which leads to a greater need for sedation. Thus, it is essential to appropriately tailor and accurately perform noninvasive imaging tests in these younger patients. This article is an overview of updated imaging techniques used to assess liver disease quantitatively in children. The common initial imaging study for diffuse liver disease in pediatric patients is ultrasound. In addition to preexisting echo analysis, newly developed attenuation imaging techniques have been introduced to evaluate fatty liver. Ultrasound elastography is also now actively used to evaluate liver conditions, and the broad age spectrum of the pediatric population requires caution to be taken even in the selection of probes. Magnetic resonance imaging (MRI) is another important imaging tool used to evaluate liver disease despite requiring sedation or anesthesia in young children because it allows quantitative analysis with sequences such as fat analysis and MR elastography. In addition to ultrasound and MRI, we review quantitative imaging methods specifically for fatty liver, Wilson disease, biliary atresia, hepatic fibrosis, Fontan-associated liver disease, autoimmune hepatitis, sinusoidal obstruction syndrome, and the transplanted liver. Lastly, concerns such as growth and motion that need to be addressed specifically for children are summarized.

Evolving and Novel Applications of Artificial Intelligence in Thoracic Imaging

The advent of artificial intelligence (AI) is revolutionizing medicine, particularly radiology. With the development of newer models, AI applications are demonstrating improved performance and versatile utility in the clinical setting. Thoracic imaging is an area of profound interest, given the prevalence of chest imaging and the significant health implications of thoracic diseases. This review aims to highlight the promising applications of AI within thoracic imaging. It examines the role of AI, including its contributions to improving diagnostic evaluation and interpretation, enhancing workflow, and aiding in invasive procedures. Next, it further highlights the current challenges and limitations faced by AI, such as the necessity of 'big data', ethical and legal considerations, and bias in representation. Lastly, it explores the potential directions for the application of AI in thoracic radiology.

Hybrid transformer convolutional neural network-based radiomics models for osteoporosis screening in routine CT

OBJECTIVE

Early diagnosis of osteoporosis is crucial to prevent osteoporotic vertebral fracture and complications of spine surgery. We aimed to conduct a hybrid transformer convolutional neural network (HTCNN)-based radiomics model for osteoporosis screening in routine CT.

METHODS

To investigate the HTCNN algorithm for vertebrae and trabecular segmentation, 92 training subjects and 45 test subjects were employed. Furthermore, we included 283 vertebral bodies and randomly divided them into the training cohort (n = 204) and test cohort (n = 79) for radiomics analysis. Area receiver operating characteristic curves (AUCs) and decision curve analysis (DCA) were applied to compare the performance and clinical value between radiomics models and Hounsfield Unit (HU) values to detect dual-energy X-ray absorptiometry (DXA) based osteoporosis.

RESULTS

HTCNN algorithm revealed high precision for the segmentation of the vertebral body and trabecular compartment. In test sets, the mean dice scores reach 0.968 and 0.961. 12 features from the trabecular compartment and 15 features from the entire vertebral body were used to calculate the radiomics score (rad score). Compared with HU values and trabecular rad-score, the vertebrae rad-score suggested the best efficacy for osteoporosis and non-osteoporosis discrimination (training group: AUC = 0.95, 95%CI 0.91-0.99; test group: AUC = 0.97, 95%CI 0.93-1.00) and the differences were significant in test group according to the DeLong test (p < 0.05).

CONCLUSIONS

This retrospective study demonstrated the superiority of the HTCNN-based vertebrae radiomics model for osteoporosis discrimination in routine CT.

Improving radiomics reproducibility using deep learning-based image conversion of CT reconstruction algorithms in hepatocellular carcinoma patients

OBJECTIVES

CT reconstruction algorithms affect radiomics reproducibility. In this study, we evaluate the effect of deep learning-based image conversion on CT reconstruction algorithms.

METHODS

This study included 78 hepatocellular carcinoma (HCC) patients who underwent four-phase liver CTs comprising non-contrast, late arterial (LAP), portal venous (PVP), and delayed phase (DP), reconstructed using both filtered back projection (FBP) and advanced modeled iterative reconstruction (ADMIRE). PVP images were used to train a convolutional neural network (CNN) model to convert images from FBP to ADMIRE and vice versa. LAP, PVP, and DP images were used for validation and testing. Radiomic features were extracted for each patient with a semi-automatic segmentation tool. We used concordance correlation coefficients (CCCs) to evaluate the radiomics reproducibility for original FBP (oFBP) vs. original ADMIRE (oADMIRE), oFBP vs. converted FBP (cFBP), and oADMIRE vs. converted ADMIRE (cADMIRE).

RESULTS

In the test group including 30 patients, the CCC and proportion of reproducible features (CCC ≥ 0.85) for oFBP vs. oADMIRE were 0.65 and 32.9% (524/1595) for LAP, 0.65 and 35.9% (573/1595) for PVP, and 0.69 and 43.8% (699/1595) for DP. For oFBP vs. cFBP, the values increased to 0.92 and 83.9% (1339/1595) for LAP, 0.89 and 71.0% (1133/1595) for PVP, and 0.90 and 79.7% (1271/1595) for DP. Similarly, for oADMIRE vs. cADMIRE, the values increased to 0.87 and 68.1% (1086/1595) for LAP, 0.91 and 82.1% (1309/1595) for PVP, and 0.89 and 76.2% (1216/1595) for DP.

CONCLUSIONS

CNN-based image conversion between CT reconstruction algorithms improved the radiomics reproducibility of HCCs.

CLINICAL RELEVANCE STATEMENT

This study demonstrates that using a CNN-based image conversion technique significantly improves the reproducibility of radiomic features in HCCs, highlighting its potential for enhancing radiomics research in HCC patients.

KEY POINTS

Radiomics reproducibility of HCC was improved via CNN-based image conversion between two different CT reconstruction algorithms. This is the first clinical study to demonstrate improvements across a range of radiomic features in HCC patients. This study promotes the reproducibility and generalizability of different CT reconstruction algorithms in radiomics research.

Impacts of Adaptive Statistical Iterative Reconstruction-V and Deep Learning Image Reconstruction Algorithms on Robustness of CT Radiomics Features: Opportunity for Minimizing Radiomics Variability Among Scans of Different Dose Levels

This study aims to investigate the influence of adaptive statistical iterative reconstruction-V (ASIR-V) and deep learning image reconstruction (DLIR) on CT radiomics feature robustness. A standardized phantom was scanned under single-energy CT (SECT) and dual-energy CT (DECT) modes at standard and low (20 and 10 mGy) dose levels. Images of SECT 120 kVp and corresponding DECT 120 kVp-like virtual monochromatic images were generated with filtered back-projection (FBP), ASIR-V at 40% (AV-40) and 100% (AV-100) blending levels, and DLIR algorithm at low (DLIR-L), medium (DLIR-M), and high (DLIR-H) strength levels. Ninety-four features were extracted via Pyradiomics. Reproducibility of features was calculated between standard and low dose levels, between reconstruction algorithms in reference to FBP images, and within scan mode, using intraclass correlation coefficient (ICC) and concordance correlation coefficient (CCC). The average percentage of features with ICC > 0.90 and CCC > 0.90 between the two dose levels was 21.28% and 20.75% in AV-40 images, and 39.90% and 35.11% in AV-100 images, respectively, and increased from 15.43 to 45.22% and from 15.43 to 44.15% with an increasing strength level of DLIR. The average percentage of features with ICC > 0.90 and CCC > 0.90 in reference to FBP images was 26.07% and 25.80% in AV-40 images, and 18.88% and 18.62% in AV-100 images, respectively, and decreased from 27.93 to 17.82% and from 27.66 to 17.29% with an increasing strength level of DLIR. DLIR and ASIR-V algorithms showed low reproducibility in reference to FBP images, while the high-strength DLIR algorithm provides an opportunity for minimizing radiomics variability due to dose reduction.

Imaging Evaluation of Peritoneal Metastasis: Current and Promising Techniques

Early diagnosis, accurate assessment, and localization of peritoneal metastasis (PM) are essential for the selection of appropriate treatments and surgical guidance. However, available imaging modalities (computed tomography [CT], conventional magnetic resonance imaging [MRI], and 18fluorodeoxyglucose positron emission tomography [PET]/CT) have limitations. The advent of new imaging techniques and novel molecular imaging agents have revealed molecular processes in the tumor microenvironment as an application for the early diagnosis and assessment of PM as well as real-time guided surgical resection, which has changed clinical management. In contrast to clinical imaging, which is purely qualitative and subjective for interpreting macroscopic structures, radiomics and artificial intelligence (AI) capitalize on high-dimensional numerical data from images that may reflect tumor pathophysiology. A predictive model can be used to predict the occurrence, recurrence, and prognosis of PM, thereby avoiding unnecessary exploratory surgeries. This review summarizes the role and status of different imaging techniques, especially new imaging strategies such as spectral photon-counting CT, fibroblast activation protein inhibitor (FAPI) PET/CT, near-infrared fluorescence imaging, and PET/MRI, for early diagnosis, assessment of surgical indications, and recurrence monitoring in patients with PM. The clinical applications, limitations, and solutions for fluorescence imaging, radiomics, and AI are also discussed.

Differentiation of benign and malignant vertebral fractures using a convolutional neural network to extract CT-based texture features

PURPOSE

To assess the diagnostic performance of three-dimensional (3D) CT-based texture features (TFs) using a convolutional neural network (CNN)-based framework to differentiate benign (osteoporotic) and malignant vertebral fractures (VFs).

METHODS

A total of 409 patients who underwent routine thoracolumbar spine CT at two institutions were included. VFs were categorized as benign or malignant using either biopsy or imaging follow-up of at least three months as standard of reference. Automated detection, labelling, and segmentation of the vertebrae were performed using a CNN-based framework ( https://anduin.bonescreen.de ). Eight TFs were extracted: Varianceglobal, Skewnessglobal, energy, entropy, short-run emphasis (SRE), long-run emphasis (LRE), run-length non-uniformity (RLN), and run percentage (RP). Multivariate regression models adjusted for age and sex were used to compare TFs between benign and malignant VFs.

RESULTS

Skewnessglobal showed a significant difference between the two groups when analyzing fractured vertebrae from T1 to L6 (benign fracture group: 0.70 [0.64-0.76]; malignant fracture group: 0.59 [0.56-0.63]; and p = 0.017), suggesting a higher skewness in benign VFs compared to malignant VFs.

CONCLUSION

Three-dimensional CT-based global TF skewness assessed using a CNN-based framework showed significant difference between benign and malignant thoracolumbar VFs and may therefore contribute to the clinical diagnostic work-up of patients with VFs.

Deep learning imaging reconstruction of reduced-dose 40 keV virtual monoenergetic imaging for early detection of colorectal cancer liver metastases

OBJECTIVE

To explore whether reduced-dose (RD) gemstone spectral imaging (GSI) and deep learning image reconstruction (DLIR) of 40 keV virtual monoenergetic image (VMI) enhanced the early detection and diagnosis of colorectal cancer liver metastases (CRLM).

METHODS

Thirty-five participants with pathologically confirmed colorectal cancer were prospectively enrolled from March to August 2022 after routine care abdominal computed tomography (CT). GSI mode was used for contrast-enhanced CT, and two portal venous phase CT images were obtained [standard-dose (SD) CT dose index (CTDIvol) = 15.51 mGy, RD CTDIvol = 7.95 mGy]. The 40 keV-VMI were reconstructed via filtered back projection (FBP) and iterative reconstruction (ASIR-V 60 %, AV60) of both SD and RD images. RD medium-strength deep learning image reconstruction (DLIR-M) and RD high-strength deep learning image reconstruction (DLIR-H) were used to reconstruct the 40 keV-VMI. The contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR) of the liver and the lesions were objectively evaluated. The overall image quality, lesion conspicuity, and diagnostic confidence were subjectively evaluated, to compare the differences in evaluation results among the different images.

RESULTS

All 35 participants (mean age: 59.51 ± 11.01 years; 14 females) underwent SD and RD GSI portal venous-phase CT scans. The dose-length product of the RD GSI scan was reduced by 49-53 % lower than that of the SD GSI scan (420.22 ± 31.95) vs (817.58 ± 60.56). A total of 219 lesions were identified, including 55 benign lesions and 164 metastases, with an average size of 7.37 ± 4.14 mm. SD-FBP detected 207 lesions, SD-AV60 detected 201 lesions, and DLIR-M and DLIR-H detected 199 and 190 lesions, respectively. For lesions ≤ 5 mm, there was no statistical difference between SD-FBP vs DLIR-M (χ2McNemar = 1.00, P = 0.32) and SD-AV60 vs DLIR-M (χ2McNemar = 0.33, P = 0.56) in the detection rate. The CNR, SNR, and noise of DLIR-M and DLIR-H 40 keV-VMI images were better than those of SD-FBP images (P < 0.01) but did not differ significantly from those of SD-AV60 images (P > 0.05). When the lesions ≤ 5 mm, there were statistical differences in the overall diagnostic sensitivity of lesions compared with SD-FBP, SD-AV60, DLIR-M and DLIR-H (P＜0.01). There were no statistical differences in the sensitivity of lesions diagnosis between SD-FBP, SD-AV60 and DLIR-M (both P＞0.05). However, the DLIR-M subjective image quality and lesion diagnostic confidence were higher for SD-FBP (both P < 0.01).

CONCLUSION

Reduced dose DLIR-M of 40 keV-VMI can be used for routine follow-up care of colorectal cancer patients, to optimize evaluations and ensure CT image quality. Meanwhile, the detection rate and diagnostic sensitivity and specificity of small lesions, early liver metastases is not obviously reduced.

Artificial intelligence-based radiomics in bone tumors: Technical advances and clinical application

Radiomics is the extraction of predefined mathematic features from medical images for predicting variables of clinical interest. Recent research has demonstrated that radiomics can be processed by artificial intelligence algorithms to reveal complex patterns and trends for diagnosis, and prediction of prognosis and response to treatment modalities in various types of cancer. Artificial intelligence tools can utilize radiological images to solve next-generation issues in clinical decision making. Bone tumors can be classified as primary and secondary (metastatic) tumors. Osteosarcoma, Ewing sarcoma, and chondrosarcoma are the dominating primary tumors of bone. The development of bone tumor model systems and relevant research, and the assessment of novel treatment methods are ongoing to improve clinical outcomes, notably for patients with metastases. Artificial intelligence and radiomics have been utilized in almost full spectrum of clinical care of bone tumors. Radiomics models have achieved excellent performance in the diagnosis and grading of bone tumors. Furthermore, the models enable to predict overall survival, metastases, and recurrence. Radiomics features have exhibited promise in assisting therapeutic planning and evaluation, especially neoadjuvant chemotherapy. This review provides an overview of the evolution and opportunities for artificial intelligence in imaging, with a focus on hand-crafted features and deep learning-based radiomics approaches. We summarize the current application of artificial intelligence-based radiomics both in primary and metastatic bone tumors, and discuss the limitations and future opportunities of artificial intelligence-based radiomics in this field. In the era of personalized medicine, our in-depth understanding of emerging artificial intelligence-based radiomics approaches will bring innovative solutions to bone tumors and achieve clinical application.

Deep learning-assisted diagnosis of benign and malignant parotid tumors based on contrast-enhanced CT: a multicenter study

OBJECTIVES

To develop deep learning-assisted diagnosis models based on CT images to facilitate radiologists in differentiating benign and malignant parotid tumors.

METHODS

Data from 573 patients with histopathologically confirmed parotid tumors from center 1 (training set: n = 269; internal-testing set: n = 116) and center 2 (external-testing set: n = 188) were retrospectively collected. Six deep learning models (MobileNet V3, ShuffleNet V2, Inception V3, DenseNet 121, ResNet 50, and VGG 19) based on arterial-phase CT images, and a baseline support vector machine (SVM) model integrating clinical-radiological features with handcrafted radiomics signatures were constructed. The performance of senior and junior radiologists with and without optimal model assistance was compared. The net reclassification index (NRI) and integrated discrimination improvement (IDI) were calculated to evaluate the clinical benefit of using the optimal model.

RESULTS

MobileNet V3 had the best predictive performance, with sensitivity increases of 0.111 and 0.207 (p < 0.05) in the internal- and external-testing sets, respectively, relative to the SVM model. Clinical benefit and overall efficiency of junior radiologist were significantly improved with model assistance; for the internal- and external-testing sets, respectively, the AUCs improved by 0.128 and 0.102 (p < 0.05), the sensitivity improved by 0.194 and 0.120 (p < 0.05), the NRIs were 0.257 and 0.205 (p < 0.001), and the IDIs were 0.316 and 0.252 (p < 0.001).

CONCLUSIONS

The developed deep learning models can assist radiologists in achieving higher diagnostic performance and hopefully provide more valuable information for clinical decision-making in patients with parotid tumors.

KEY POINTS

• The developed deep learning models outperformed the traditional SVM model in predicting benign and malignant parotid tumors. • Junior radiologist can obtain greater clinical benefits with assistance from the optimal deep learning model. • The clinical decision-making process can be accelerated in patients with parotid tumors using the established deep learning model.

Generative Adversarial Network-Based Image Conversion Among Different Computed Tomography Protocols and Vendors: Effects on Accuracy and Variability in Quantifying Regional Disease Patterns of Interstitial Lung Disease

OBJECTIVE

To assess whether computed tomography (CT) conversion across different scan parameters and manufacturers using a routable generative adversarial network (RouteGAN) can improve the accuracy and variability in quantifying interstitial lung disease (ILD) using a deep learning-based automated software.

MATERIALS AND METHODS

This study included patients with ILD who underwent thin-section CT. Unmatched CT images obtained using scanners from four manufacturers (vendors A-D), standard- or low-radiation doses, and sharp or medium kernels were classified into groups 1-7 according to acquisition conditions. CT images in groups 2-7 were converted into the target CT style (Group 1: vendor A, standard dose, and sharp kernel) using a RouteGAN. ILD was quantified on original and converted CT images using a deep learning-based software (Aview, Coreline Soft). The accuracy of quantification was analyzed using the dice similarity coefficient (DSC) and pixel-wise overlap accuracy metrics against manual quantification by a radiologist. Five radiologists evaluated quantification accuracy using a 10-point visual scoring system.

RESULTS

Three hundred and fifty CT slices from 150 patients (mean age: 67.6 ± 10.7 years; 56 females) were included. The overlap accuracies for quantifying total abnormalities in groups 2-7 improved after CT conversion (original vs. converted: 0.63 vs. 0.68 for DSC, 0.66 vs. 0.70 for pixel-wise recall, and 0.68 vs. 0.73 for pixel-wise precision; P < 0.002 for all). The DSCs of fibrosis score, honeycombing, and reticulation significantly increased after CT conversion (0.32 vs. 0.64, 0.19 vs. 0.47, and 0.23 vs. 0.54, P < 0.002 for all), whereas those of ground-glass opacity, consolidation, and emphysema did not change significantly or decreased slightly. The radiologists' scores were significantly higher (P < 0.001) and less variable on converted CT.

CONCLUSION

CT conversion using a RouteGAN can improve the accuracy and variability of CT images obtained using different scan parameters and manufacturers in deep learning-based quantification of ILD.

Deep Learning-Based Computed Tomography Image Standardization to Improve Generalizability of Deep Learning-Based Hepatic Segmentation

OBJECTIVE

We aimed to investigate whether image standardization using deep learning-based computed tomography (CT) image conversion would improve the performance of deep learning-based automated hepatic segmentation across various reconstruction methods.

MATERIALS AND METHODS

We collected contrast-enhanced dual-energy CT of the abdomen that was obtained using various reconstruction methods, including filtered back projection, iterative reconstruction, optimum contrast, and monoenergetic images with 40, 60, and 80 keV. A deep learning based image conversion algorithm was developed to standardize the CT images using 142 CT examinations (128 for training and 14 for tuning). A separate set of 43 CT examinations from 42 patients (mean age, 10.1 years) was used as the test data. A commercial software program (MEDIP PRO v2.0.0.0, MEDICALIP Co. Ltd.) based on 2D U-NET was used to create liver segmentation masks with liver volume. The original 80 keV images were used as the ground truth. We used the paired t-test to compare the segmentation performance in the Dice similarity coefficient (DSC) and difference ratio of the liver volume relative to the ground truth volume before and after image standardization. The concordance correlation coefficient (CCC) was used to assess the agreement between the segmented liver volume and ground-truth volume.

RESULTS

The original CT images showed variable and poor segmentation performances. The standardized images achieved significantly higher DSCs for liver segmentation than the original images (DSC [original, 5.40%-91.27%] vs. [standardized, 93.16%-96.74%], all P < 0.001). The difference ratio of liver volume also decreased significantly after image conversion (original, 9.84%-91.37% vs. standardized, 1.99%-4.41%). In all protocols, CCCs improved after image conversion (original, -0.006-0.964 vs. standardized, 0.990-0.998).

CONCLUSION

Deep learning-based CT image standardization can improve the performance of automated hepatic segmentation using CT images reconstructed using various methods. Deep learning-based CT image conversion may have the potential to improve the generalizability of the segmentation network.

Influence of Image Processing on Radiomic Features From Magnetic Resonance Imaging

OBJECTIVE

Before implementing radiomics in routine clinical practice, comprehensive knowledge about the repeatability and reproducibility of radiomic features is required. The aim of this study was to systematically investigate the influence of image processing parameters on radiomic features from magnetic resonance imaging (MRI) in terms of feature values as well as test-retest repeatability.

MATERIALS AND METHODS

Utilizing a phantom consisting of 4 onions, 4 limes, 4 kiwifruits, and 4 apples, we acquired a test-retest dataset featuring 3 of the most commonly used MRI sequences on a 3 T scanner, namely, a T1-weighted, a T2-weighted, and a fluid-attenuated inversion recovery sequence, each at high and low resolution. After semiautomatic image segmentation, image processing with systematic variation of image processing parameters was performed, including spatial resampling, intensity discretization, and intensity rescaling. For each respective image processing setting, a total of 45 radiomic features were extracted, corresponding to the following 7 matrices/feature classes: conventional indices, histogram matrix, shape matrix, gray-level zone length matrix, gray-level run length matrix, neighboring gray-level dependence matrix, and gray-level cooccurrence matrix. Systematic differences of individual features between different resampling steps were assessed using 1-way analysis of variance with Tukey-type post hoc comparisons to adjust for multiple testing. Test-retest repeatability of radiomic features was measured using the concordance correlation coefficient, dynamic range, and intraclass correlation coefficient.

RESULTS

Image processing influenced radiological feature values. Regardless of the acquired sequence and feature class, significant differences ( P < 0.05) in feature values were found when the size of the resampled voxels was too large, that is, bigger than 3 mm. Almost all higher-order features depended strongly on intensity discretization. The effects of intensity rescaling were negligible except for some features derived from T1-weighted sequences. For all sequences, the percentage of repeatable features (concordance correlation coefficient and dynamic range ≥ 0.9) varied considerably depending on the image processing settings. The optimal image processing setting to achieve the highest percentage of stable features varied per sequence. Irrespective of image processing, the fluid-attenuated inversion recovery sequence in high-resolution overall yielded the highest number of stable features in comparison with the other sequences (89% vs 64%-78% for the respective optimal image processing settings). Across all sequences, the most repeatable features were generally obtained for a spatial resampling close to the originally acquired voxel size and an intensity discretization to at least 32 bins.

CONCLUSION

Variation of image processing parameters has a significant impact on the values of radiomic features as well as their repeatability. Furthermore, the optimal image processing parameters differ for each MRI sequence. Therefore, it is recommended that these processing parameters be determined in corresponding test-retest scans before clinical application. Extensive repeatability, reproducibility, and validation studies as well as standardization are required before quantitative image analysis and radiomics can be reliably translated into routine clinical care.

Deep learning-based reconstruction on cardiac CT yields distinct radiomic features compared to iterative and filtered back projection reconstructions

We aimed to determine the effects of deep learning-based reconstruction (DLR) on radiomic features obtained from cardiac computed tomography (CT) by comparing with iterative reconstruction (IR), and filtered back projection (FBP). A total of 284 consecutive patients with 285 cardiac CT scans that were reconstructed with DLR, IR, and FBP, were retrospectively enrolled. Radiomic features were extracted from the left ventricular (LV) myocardium, and from the periprosthetic mass if patients had cardiac valve replacement. Radiomic features of LV myocardium from each reconstruction were compared using a fitting linear mixed model. Radiomics models were developed to diagnose periprosthetic abnormality, and the performance was evaluated using the area under the receiver characteristics curve (AUC). Most radiomic features of LV myocardium (73 of 88) were significantly different in pairwise comparisons between all three reconstruction methods (P < 0.05). The radiomics model on IR exhibited the best diagnostic performance (AUC 0.948, 95% CI 0.880-1), relative to DLR (AUC 0.873, 95% CI 0.735-1) and FBP (AUC 0.875, 95% CI 0.731-1), but these differences did not reach significance (P > 0.05). In conclusion, applying DLR to cardiac CT scans yields radiomic features distinct from those obtained with IR and FBP, implying that feature robustness is not guaranteed when applying DLR.

Deep learning reconstruction improves radiomics feature stability and discriminative power in abdominal CT imaging: a phantom study

OBJECTIVES

To compare image quality of deep learning reconstruction (AiCE) for radiomics feature extraction with filtered back projection (FBP), hybrid iterative reconstruction (AIDR 3D), and model-based iterative reconstruction (FIRST).

METHODS

Effects of image reconstruction on radiomics features were investigated using a phantom that realistically mimicked a 65-year-old patient's abdomen with hepatic metastases. The phantom was scanned at 18 doses from 0.2 to 4 mGy, with 20 repeated scans per dose. Images were reconstructed with FBP, AIDR 3D, FIRST, and AiCE. Ninety-three radiomics features were extracted from 24 regions of interest, which were evenly distributed across three tissue classes: normal liver, metastatic core, and metastatic rim. Features were analyzed in terms of their consistent characterization of tissues within the same image (intraclass correlation coefficient ≥ 0.75), discriminative power (Kruskal-Wallis test p value < 0.05), and repeatability (overall concordance correlation coefficient ≥ 0.75).

RESULTS

The median fraction of consistent features across all doses was 6%, 8%, 6%, and 22% with FBP, AIDR 3D, FIRST, and AiCE, respectively. Adequate discriminative power was achieved by 48%, 82%, 84%, and 92% of features, and 52%, 20%, 17%, and 39% of features were repeatable, respectively. Only 5% of features combined consistency, discriminative power, and repeatability with FBP, AIDR 3D, and FIRST versus 13% with AiCE at doses above 1 mGy and 17% at doses ≥ 3 mGy. AiCE was the only reconstruction technique that enabled extraction of higher-order features.

CONCLUSIONS

AiCE more than doubled the yield of radiomics features at doses typically used clinically. Inconsistent tissue characterization within CT images contributes significantly to the poor stability of radiomics features.

KEY POINTS

• Image quality of CT images reconstructed with filtered back projection and iterative methods is inadequate for the majority of radiomics features due to inconsistent tissue characterization, low discriminative power, or low repeatability. • Deep learning reconstruction enhances image quality for radiomics and more than doubled the feature yield at doses that are typically used in clinical CT imaging. • Image reconstruction algorithms can optimize image quality for more reliable quantification of tissues in CT images.