Fully automated hybrid approach to predict the IDH mutation status of gliomas via deep learning and radiomics.

BACKGROUND: Glioma prognosis depends on isocitrate dehydrogenase (IDH) mutation status. We aimed to predict the IDH status of gliomas from preoperative MR images using a fully automated hybrid approach with convolutional neural networks (CNNs) and radiomics. METHODS: We reviewed 1166 preoperative MR images of gliomas (grades II-IV) from Severance Hospital (n = 856), Seoul National University Hospital (SNUH; n = 107), and The Cancer Imaging Archive (TCIA; n = 203). The Severance set was subdivided into the development (n = 727) and internal test (n = 129) sets. Based on T1 postcontrast, T2, and fluid-attenuated inversion recovery images, a fully automated model was developed that comprised a CNN for tumor segmentation (Model 1) and CNN-based classifier for IDH status prediction (Model 2) that uses a hybrid approach based on 2D tumor images and radiomic features from 3D tumor shape and loci guided by Model 1. The trained model was tested on internal (a subset of the Severance set) and external (SNUH and TCIA) test sets. RESULTS: The CNN for tumor segmentation (Model 1) achieved a dice coefficient of 0.86-0.92 across datasets. Our hybrid model achieved accuracies of 93.8%, 87.9%, and 78.8%, with areas under the receiver operating characteristic curves of 0.96, 0.94, and 0.86 and areas under the precision-recall curves of 0.88, 0.82, and 0.81 in the internal test, SNUH, and TCIA sets, respectively. CONCLUSIONS: Our fully automated hybrid model demonstrated the potential to be a highly reproducible and generalizable tool across different datasets for the noninvasive prediction of the IDH status of gliomas.

Squamous Cell Carcinoma and Lymphoma of the Oropharynx: Differentiation Using a Radiomics Approach.

The purpose of this study was to evaluate the diagnostic performance of magnetic resonance (MR) radiomics-based machine learning algorithms in differentiating squamous cell carcinoma (SCC) from lymphoma in the oropharynx. MR images from 87 patients with oropharyngeal SCC (n=68) and lymphoma (n=19) were reviewed retrospectively. Tumors were semi-automatically segmented on contrast-enhanced T1-weighted images registered to T2-weighted images, and radiomic features (n=202) were extracted from contrast-enhanced T1- and T2-weighted images. The radiomics classifier was built using elastic-net regularized generalized linear model analyses with nested five-fold cross-validation. The diagnostic abilities of the radiomics classifier and visual assessment by two head and neck radiologists were evaluated using receiver operating characteristic (ROC) analyses for distinguishing SCC from lymphoma. Nineteen radiomics features were selected at least twice during the five-fold cross-validation. The mean area under the ROC curve (AUC) of the radiomics classifier was 0.750 [95% confidence interval (CI), 0.613-0.887], with a sensitivity of 84.2%, specificity of 60.3%, and an accuracy of 65.5%. Two human readers yielded AUCs of 0.613 (95% CI, 0.467-0.759) and 0.663 (95% CI, 0.531-0.795), respectively. The radiomics-based machine learning model can be useful for differentiating SCC from lymphoma of the oropharynx.

Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation.

We evaluated the diagnostic performance and generalizability of traditional machine learning and deep learning models for distinguishing glioblastoma from single brain metastasis using radiomics. The training and external validation cohorts comprised 166 (109 glioblastomas and 57 metastases) and 82 (50 glioblastomas and 32 metastases) patients, respectively. Two-hundred-and-sixty-five radiomic features were extracted from semiautomatically segmented regions on contrast-enhancing and peritumoral T2 hyperintense masks and used as input data. For each of a deep neural network (DNN) and seven traditional machine learning classifiers combined with one of five feature selection methods, hyperparameters were optimized through tenfold cross-validation in the training cohort. The diagnostic performance of the optimized models and two neuroradiologists was tested in the validation cohort for distinguishing glioblastoma from metastasis. In the external validation, DNN showed the highest diagnostic performance, with an area under receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy of 0.956 (95% confidence interval [CI], 0.918-0.990), 90.6% (95% CI, 80.5-100), 88.0% (95% CI, 79.0-97.0), and 89.0% (95% CI, 82.3-95.8), respectively, compared to the best-performing traditional machine learning model (adaptive boosting combined with tree-based feature selection; AUC, 0.890 (95% CI, 0.823-0.947)) and human readers (AUC, 0.774 [95% CI, 0.685-0.852] and 0.904 [95% CI, 0.852-0.951]). The results demonstrated deep learning using radiomic features can be useful for distinguishing glioblastoma from metastasis with good generalizability.

A deep learning algorithm may automate intracranial aneurysm detection on MR angiography with high diagnostic performance.

OBJECTIVES: To develop a deep learning algorithm for automated detection and localization of intracranial aneurysms on time-of-flight MR angiography and evaluate its diagnostic performance. METHODS: In a retrospective and multicenter study, MR images with aneurysms based on radiological reports were extracted. The examinations were randomly divided into two data sets: training set of 468 examinations and internal test set of 120 examinations. Additionally, 50 examinations without aneurysms were randomly selected and added to the internal test set. External test data set consisted of 56 examinations with intracranial aneurysms and 50 examinations without aneurysms, which were extracted based on radiological reports from a different institution. After manual ground truth segmentation of aneurysms, a deep learning algorithm based on 3D ResNet architecture was established with the training set. Its sensitivity, positive predictive value, and specificity were evaluated in the internal and external test sets. RESULTS: MR images included 551 aneurysms (mean diameter, 4.17 ± 2.49 mm) in the training, 147 aneurysms (mean diameter, 3.98 ± 2.11 mm) in the internal test, 63 aneurysms (mean diameter, 3.23 ± 1.69 mm) in the external test sets. The sensitivity, the positive predictive value, and the specificity were 87.1%, 92.8%, and 92.0% for the internal test set and 85.7%, 91.5%, and 98.0% for the external test set, respectively. CONCLUSION: A deep learning algorithm detected intracranial aneurysms with a high diagnostic performance which was validated using external data set. KEY POINTS: • A deep learning-based algorithm for the automated diagnosis of intracranial aneurysms demonstrated a high sensitivity, positive predictive value, and specificity. • The high diagnostic performance of the algorithm was validated using external test data set from a different institution with a different scanner. • The algorithm might be robust and effective for general use in real clinical settings.

The added prognostic value of radiological phenotype combined with clinical features and molecular subtype in anaplastic gliomas.

PURPOSE: To determine whether radiological phenotype can improve the predictive performance of the risk model based on molecular subtype and clinical risk factors in anaplastic glioma patients. METHODS: This retrospective study was approved by our institutional review board with waiver of informed consent. MR images of 86 patients with pathologically diagnosed anaplastic glioma (WHO grade III) between January 2007 and February 2016 were analyzed according to the Visually Accessible Rembrandt Images (VASARI) features set. Significant imaging findings were selected to generate a radiological risk score (RRS) for overall survival (OS) and progression-free survival (PFS) using the least absolute shrinkage and selection operator (LASSO) Cox regression model. The prognostic value of RRS was evaluated with multivariate Cox regression including molecular subtype and clinical risk factors. The C-indices of multivariate models with and without RRS were compared by bootstrapping. RESULTS: Eight VASARI features contributed to RRS for OS and six contributed to PFS. Multifocality or multicentricity was the most influential feature, followed by restricted diffusion. RRS was significantly associated with OS and PFS (P < .001), as well as age and molecular subtype. The multivariate model with RRS demonstrated a significantly higher predictive performance than the model without (C-index difference: 0.074, 95% confidence interval [CI]: 0.031, 0.148 for OS; C-index difference: 0.054, 95% CI: 0.014, 0.123 for PFS). CONCLUSION: RRS derived from VASARI features was an independent predictor of survival in patients with anaplastic gliomas. The addition of RRS significantly improved the predictive performance of the molecular feature based model.

Radiomic MRI Phenotyping of Glioblastoma: Improving Survival Prediction.

Purpose To investigate whether radiomic features at MRI improve survival prediction in patients with glioblastoma multiforme (GBM) when they are integrated with clinical and genetic profiles. Materials and Methods Data in patients with a diagnosis of GBM between December 2009 and January 2017 (217 patients) were retrospectively reviewed up to May 2017 and allocated to training and test sets (3:1 ratio). Radiomic features (n = 796) were extracted from multiparametric MRI. A random survival forest (RSF) model was trained with the radiomic features along with clinical and genetic profiles (O-6-methylguanine-DNA-methyltransferase promoter methylation and isocitrate dehydrogenase 1 mutation statuses) to predict overall survival (OS) and progression-free survival (PFS). The RSF models were validated on the test set. The incremental values of radiomic features were evaluated by using the integrated area under the receiver operating characteristic curve (iAUC). Results The 217 patients had a mean age of 57.9 years, and there were 87 female patients (age range, 22-81 years) and 130 male patients (age range, 17-85 years). The median OS and PFS of patients were 352 days (range, 20-1809 days) and 264 days (range, 21-1809 days), respectively. The RSF radiomics models were successfully validated on the test set (iAUC, 0.652 [95% confidence interval {CI}, 0.524, 0.769] and 0.590 [95% CI: 0.502, 0.689] for OS and PFS, respectively). The addition of a radiomics model to clinical and genetic profiles improved survival prediction when compared with models containing clinical and genetic profiles alone (P = .04 and .03 for OS and PFS, respectively). Conclusion Radiomic MRI phenotyping can improve survival prediction when integrated with clinical and genetic profiles and thus has potential as a practical imaging biomarker. © RSNA, 2018 Online supplemental material is available for this article. See also the editorial by Jain and Lui in this issue.

Primary central nervous system lymphoma and atypical glioblastoma: Differentiation using radiomics approach.

OBJECTIVES: To evaluate the diagnostic performance of magnetic resonance (MR) radiomics-based machine-learning algorithms in differentiating primary central nervous system lymphoma (PCNSL) from non-necrotic atypical glioblastoma (GBM). METHODS: Seventy-seven patients (54 individuals with PCNSL and 23 with non-necrotic atypical GBM), diagnosed from January 2009 to April 2017, were enrolled in this retrospective study. A total of 6,366 radiomics features, including shape, volume, first-order, texture, and wavelet-transformed features, were extracted from multi-parametric (post-contrast T1- and T2-weighted, and fluid attenuation inversion recovery images) and multiregional (enhanced and non-enhanced) tumour volumes. These features were subjected to recursive feature elimination and random forest (RF) analysis with nested cross-validation. The diagnostic abilities of a radiomics machine-learning classifier, apparent diffusion coefficient (ADC), and three readers, who independently classified the tumours based on conventional MR sequences, were evaluated using receiver operating characteristic (ROC) analysis. Areas under the ROC curves (AUC) of the radiomics classifier, ADC value, and the radiologists were compared. RESULTS: The mean AUC of the radiomics classifier was 0.921 (95 % CI 0.825-0.990). The AUCs of the three readers and ADC were 0.707 (95 % CI 0.622-0.793), 0.759 (95 %CI 0.656-0.861), 0.695 (95 % CI 0.590-0.800) and 0.684 (95 % CI0.560-0.809), respectively. The AUC of the radiomics-based classifier was significantly higher than those of the three readers and ADC (p< 0.001 for all). CONCLUSIONS: Large-scale radiomics with a machine-learning algorithm can be useful for differentiating PCNSL from atypical GBM, and yields a better diagnostic performance than human radiologists and ADC values. KEY POINTS: • Machine-learning algorithm radiomics can help to differentiate primary central PCNSL from GBM. • This approach yields a higher diagnostic accuracy than visual analysis by radiologists. • Radiomics can strengthen radiologists' diagnostic decisions whenever conventional MRI sequences are available.

MR lymphography for sentinel lymph node detection in patients with oral cavity cancer: Preliminary clinical study.

BACKGROUND: The purpose of this study was to evaluate the feasibility of MR lymphography with interstitial injection of a gadolinium-based contrast agent for identifying sentinel lymph nodes in patients with oral cavity cancer and clinically negative neck. METHODS: Pretreatment MR lymphography with a differential subsampling with cartesian ordering (DISCO) sequence was performed in 26 patients with resectable oral cavity cancer and clinically negative neck, after peritumoral injection of 1-mL diluted gadobutrol. The accuracy of sentinel lymph node identification by MR lymphography was assessed and compared with the final histopathological results. RESULTS: The MR lymphography consistently visualized the 44 sentinel lymph nodes in all 26 patients. In all but 1 patient with pathologically positive neck, assumed sentinel lymph nodes revealed metastatic involvement. CONCLUSION: Pretreatment MR lymphography is a safe and feasible imaging technique that can help clinicians identify sentinel lymph nodes with a high risk of occult metastases in patients with oral cavity cancer, enabling focused preoperative biopsy in these high-risk patients.

Intra-Suprasellar Schwannoma Presumably Originating from the Internal Carotid Artery Wall : Case Report and Review of the Literature.

Intracranial schwannomas are common but intra-suprasellar schwannomas are extremely rare. We represent the case of a 57-year-old woman with an intra-suprasellar tumor that was presurgically diagnosed as an anterior clinoidal meningioma. Intraoperatively it adhered tightly to the internal carotid artery wall and could not easily be dissected, possibly leading to profuse hemorrhage and sacrifice of the carotid artery with coil embolization. Histopathology demonstrated a schwannoma. We reviewed and summarized the clinical, imaging, and intraoperative findings of previously reported intra-suprasellar schwannomas.

Gadolinium deposition in the brain: association with various GBCAs using a generalized additive model.

OBJECTIVES: To determine the relationship between the number of administrations of various gadolinium-based contrast agents (GBCAs) and increased T1 signal intensity in the globus pallidus (GP) and dentate nucleus (DN). METHODS: This retrospective study included 122 patients who underwent double-dose GBCA-enhanced magnetic resonance imaging. Two radiologists calculated GP-to-thalamus (TH) signal intensity ratio, DN-to-pons signal intensity ratio and relative change (Rchange) between the baseline and final examinations. Interobserver agreement was evaluated. The relationships between Rchange and several factors, including number of each GBCA administrations, were analysed using a generalized additive model. RESULTS: Six patients (4.9%) received linear GBCAs (mean 20.8 number of administration; range 15-30), 44 patients (36.1%) received macrocyclic GBCAs (mean 26.1; range 14-51) and 72 patients (59.0%) received both types of GBCAs (mean 31.5; range 12-65). Interobserver agreement was almost perfect (0.99; 95% CI: 0.99-0.99). Rchange (DN:pons) was associated with gadodiamide (p = 0.006) and gadopentetate dimeglumine (p < 0.001), but not with other GBCAs. Rchange (GP:TH) was not associated with GBCA administration. CONCLUSIONS: Previous administration of linear agents gadoiamide and gadopentetate dimeglumine is associated with increased T1 signal intensity in the DN, whereas macrocyclic GBCAs do not show an association. KEY POINTS: • Certain linear GBCAs are associated with T1 signal change in the dentate nucleus. • The signal change is related to the administration number of certain linear GBCAs. • Difference in signal change may reflect differences in stability of agents.

Breast Microcalcifications: Diagnostic Outcomes According to Image-Guided Biopsy Method.

OBJECTIVE: To evaluate the diagnostic outcomes of ultrasonography-guided core needle biopsy (US-CNB), US-guided vacuum-assisted biopsy (US-VAB), and stereotactic-guided vacuum-assisted biopsy (S-VAB) for diagnosing suspicious breast microcalcification. MATERIALS AND METHODS: We retrospectively reviewed 336 cases of suspicious breast microcalcification in patients who subsequently underwent image-guided biopsy. US-CNB was performed for US-visible microcalcifications associated with a mass (n = 28), US-VAB for US-visible microcalcifications without an associated mass (n = 59), and S-VAB for mammogram-only visible lesions (n = 249). Mammographic findings, biopsy failure rate, false-negative rate, and underestimation rate were analyzed. Histological diagnoses and the Breast Imaging Reporting and Data System (BI-RADS) categories were reported. RESULTS: Biopsy failure rates for US-CNB, US-VAB, and S-VAB were 7.1% (2/28), 0% (0/59), and 2.8% (7/249), respectively. Three false-negative cases were detected for US-CNB and two for S-VAB. The rates of biopsy-diagnosed ductal carcinoma in situ that were upgraded to invasive cancer at surgery were 41.7% (5/12), 12.9% (4/31), and 8.6% (3/35) for US-CNB, US-VAB, and S-VAB, respectively. Sonographically visible lesions were more likely to be malignant (66.2% [51/77] vs. 23.2% [46/198]; p < 0.001) or of higher BI-RADS category (61.0% [47/77] vs. 22.2% [44/198]; p < 0.001) than sonographically invisible lesions. CONCLUSION: Ultrasonography-guided vacuum-assisted biopsy is more accurate than US-CNB when suspicious microcalcifications are detected on US. Calcifications with malignant pathology are significantly more visible on US than benign lesions.

Uterine Artery Embolization for Adenomyosis: Percentage of Necrosis Predicts Midterm Clinical Recurrence.

PURPOSE: To evaluate the effect of degree of necrosis after uterine artery embolization (UAE) on symptom recurrence at midterm clinical follow-up in patients with adenomyosis. MATERIALS AND METHODS: Women (N = 50) who underwent UAE for symptomatic adenomyosis were retrospectively analyzed. All patients underwent contrast-enhanced magnetic resonance (MR) imaging at baseline and 3 months after UAE and were followed clinically for at least 18 months. The type of adenomyosis was classified as focal or diffuse. The uterine volume and the percentage of necrosis after embolization were measured three-dimensionally on MR imaging. The percentage of the necrosis cutoff point for predicting recurrence was estimated. Patients were divided into 2 groups according to the cutoff point. The rate of recurrence was compared between groups, and risk factors for recurrence were identified. RESULTS: During the follow-up period (range, 18-48 mo), symptom recurrence occurred in 12 of 50 patients. A necrosis cutoff point of 34.3% was calculated to predict recurrence (area under the curve = 0.721; 95% confidence interval [CI] = 0.577-0.839; P = .004). Patients with < 34.3% necrosis (group A, n = 12) were at a significantly higher risk of recurrence than patients with > 34.3% necrosis (group B, n = 38; hazard ratio = 7.0; 95% CI = 2.2, 22.4; P = .001). Initial uterine volume and type of adenomyosis were not associated with recurrence. CONCLUSIONS: The percentage of necrosis in patients with adenomyosis after UAE may predict symptom recurrence at midterm follow-up. The cutoff percentage of necrosis required to predict symptom recurrence was 34.3% in this study.

Effects of adaptive statistical iterative reconstruction on radiation dose reduction and diagnostic accuracy of pediatric abdominal CT.

BACKGROUND: Since children are more radio-sensitive than adults, there is a need to minimize radiation exposure during CT exams. OBJECTIVE: To evaluate the effects of adaptive statistical iterative reconstruction (ASIR) on radiation dose reduction, image quality and diagnostic accuracy in pediatric abdominal CT. MATERIALS AND METHODS: We retrospectively reviewed the abdominal CT examinations of 41 children (24 boys and 17 girls; mean age: 10 years) with a low-dose radiation protocol and reconstructed with ASIR (the ASIR group). We also reviewed routine-dose abdominal CT examinations of 41 age- and sex-matched controls reconstructed with filtered-back projection (control group). Image quality was assessed objectively as noise measured in the liver, spleen and aorta, as well as subjectively by three pediatric radiologists for diagnostic acceptability using a four-point scale. Radiation dose and objective image qualities of each group were compared with the paired t-test. Diagnostic accuracy was evaluated by reviewing follow-up imaging studies and medical records in 2012 and 2013. RESULTS: There was 46.3% dose reduction of size-specific dose estimates in ASIR group (from 13.4 to 7.2 mGy) compared with the control group. Objective noise was higher in the liver, spleen and aorta of the ASIR group (P < 0.001). However, the subjective image quality was average or superior in 84-100% of studies. Only one image was subjectively rated as unacceptable by one reviewer. There was only one case with interpretational error in the control group and none in the ASIR group. CONCLUSION: Use of the ASIR technique resulted in greater than a 45% reduction in radiation dose without impairing subjective image quality or diagnostic accuracy in pediatric abdominal CT, despite increased objective image noise.