Junior Radiologists Research Articles

Radiological classification of dementia from anatomical MRI assisted by machine learning-derived maps

Background and purposeMany artificial intelligence tools are currently being developed to assist diagnosis of dementia from magnetic resonance imaging (MRI). However, these tools have so far been difficult to integrate in the clinical routine workflow. In this work, we propose a new simple way to use them and assess their utility for improving diagnostic accuracy. Materials and methodsWe studied 34 patients with early-onset Alzheimer's disease (EOAD), 49 with late-onset AD (LOAD), 39 with frontotemporal dementia (FTD) and 24 with depression from the pre-existing cohort CLIN-AD. Support vector machine (SVM) automatic classifiers using 3D T1 MRI were trained to distinguish: LOAD vs. Depression, FTD vs. LOAD, EOAD vs. Depression, EOAD vs. FTD. We extracted SVM weight maps, which are tridimensional representations of discriminant atrophy patterns used by the classifier to take its decisions and we printed posters of these maps. Four radiologists (2 senior neuroradiologists and 2 unspecialized junior radiologists) performed a visual classification of the 4 diagnostic pairs using 3D T1 MRI. Classifications were performed twice: first with standard radiological reading and then using SVM weight maps as a guide. ResultsDiagnostic performance was significantly improved by the use of the weight maps for the two junior radiologists in the case of FTD vs. EOAD. Improvement was over 10 points of diagnostic accuracy. ConclusionThis tool can improve the diagnostic accuracy of junior radiologists and could be integrated in the clinical routine workflow.

A cross-modal 3D deep learning for accurate lymph node metastasis prediction in clinical stage T1 lung adenocarcinoma

ObjectivesThe evaluation of lymph node (LN) status by radiologists based on preoperative computed tomography (CT) lacks high precision for early lung cancer patients; erroneous evaluations result in inappropriate therapeutic plans and increase the risk of complications. This study aims to develop a cross-modal 3D neural network based on CT images and prior clinical knowledge for accurate prediction of LN metastasis in clinical stage T1 lung adenocarcinoma. Patients and methodsFive hundred one lung adenocarcinoma patients with clinical stage T1 were enrolled. Data including: corresponding 3D nodule-centered patches of CT; prior clinical features; and pathological labels of LN status were obtained. We proposed a cross-modal deep learning system, which can successfully incorporate prior clinical knowledge and CT images into a 3D neural network to predict LN metastasis. We trained and validated our system with 401 cases and tested its performance with 100 cases. The result was compared with that of the logistic regression integration model, the single deep learning model without prior clinical knowledge integration, radiomics method, and manual evaluation by radiologists. ResultsThe model proposed DensePriNet achieved an AUC of 0.926, which is significantly higher than the logistic regression integration model (0.904) single deep learning model (0.880), and radiomics method (0.891). The Matthews Correlation Coefficient (MCC) of DensePriNet (0.705) was significantly higher than manual classification by one senior radiologist (0.534) and one junior radiologist (0.416), respectively. ConclusionThe performance of the single deep learning method is significantly higher than the radiomics method and the radiologists, and integration of prior clinical knowledge into the deep learning model enhance the diagnostic precision of LN status and facilitate the application of precision medicine.

Mentorship of junior radiologists in nonacademic radiology

Junior radiologists in nonacademic settings currently face multiple challenges, including clinical productivity, climbing the ladder to partnership, maintenance of certification, and administrative requirements. Mentorship of junior radiologists in nonacademic settings can ease their transition into nonacademic practice, foster comradery, increase radiologist retention, and support career growth. Strategies to implement and sustain effective mentorship in nonacademic radiology practice settings are described. SummaryStrategies to implement and sustain mentorship in nonacademic radiology practice settings are described.

Capacity for paediatric radiology in Nigeria: a survey of radiologists.

Children account for nearly half the population of Nigeria yet the capacity for effective imaging of children is unknown. In order to determine clinical resources for and challenges of paediatric radiology in Nigeria, certified radiologists at an exit examination in radiology for resident doctors, and resident doctors in radiology attending an update course, completed a semi-structured questionnaire detailing personal information, radiology training, practice and perceived priorities of paediatric radiology in Nigeria. Of 100 questionnaires, 80 were returned, completed (80%) by 46 (58%) certified radiologists, 14 (18%) senior and 20 (25%) junior radiology residents. Only 1 (2.2%) certified radiologist received dedicated albeit short training in paediatric radiology. Nine (20.0%) certified radiologists and 1 (2.9%) resident doctor target their practice to paediatric imaging. Only 2 (4.4%) certified radiologists devote at least half of their time to paediatric radiology. Forty-two (91%) of the certified radiologists are not aware of any radiologist primarily affiliated with a dedicated children's hospital or who practices in a dedicated paediatric unit of a mixed hospital. Eight (19%) radiologists work in a facility with a consultant who has some training in paediatric radiology. Surgeons (64%), neonatologists (53%) and neurologists (45%) were the other paediatric specialties usually available where the respondent works. Fourteen (4.9%), 4 (2.1%), and 1 (1.3%) ultrasound, X-ray and magnetic resonance imaging (MRI) units, respectively, were reserved for paediatric imaging while no fluoroscopy or computed tomography (CT) unit was dedicated to children. Lack of dedicated equipment (44%) and trained paediatric radiologists (24%) were the main challenges to paediatric imaging as perceived by the respondents. Substantial deficits in human and material resources require informed investment in dedicated equipment and training to boost capacity for paediatric radiology in Nigeria.

Comparison of the performance between Frontier ASPECTS software and different levels of radiologists on assessing CT examinations of acute ischaemic stroke patients

To compare the performance of Frontier Alberta Stroke Program Early CT Score (ASPECTS) software with different levels of radiologists in assessing computed tomography (CT) examinations of patients with early acute ischaemic stroke (AIS), and to evaluate whether this software can improve the performance of rating by less experienced radiologists. Unenhanced brain CT examinations of 55 patients with acute middle cerebral artery ischaemia were scored separately by Frontier, two senior radiologists, and two junior radiologists retrospectively and blinded to any clinical information. Two junior radiologists then scored again with the assist of Frontier. The reference standard was defined as the ASPECTS on Follow-up unenhanced CT scored by another two non-blinded independent experts on a consensus basis. Statistical analysis was performed using intraclass correlation coefficient (ICC) analysis and Bland-Altman plots. Frontier and senior radiologists in ASPECTS reading have excellent agreement with the reference standard (r=0.842 and 0.803, respectively), while only a good agreement was found between junior radiologists and reference standard (r=0.680). Bland-Altman analysis revealed the mean ASPECTS difference and SD difference of junior radiologists were larger (mean difference=1.35; SD=1.42) than that of Frontier and senior radiologists with reference standard (mean difference=0.16, 0.22; SD=1.24, 1.13, respectively). However, with the assist of Frontier, the agreement between junior radiologists and reference standard was improved from good (r=0.680) to excellent (r =0.852), and the mean ASPECTS difference and SD difference were reduced. High agreement in ASPECTS rating between senior radiologists, Frontier, and expert consensus reading was found. Moreover, Frontier can improve the performance of less experienced radiologists to assess the ASPECTS of patients with AIS.

Standard Diffusion-Weighted Imaging in the Brain Can Detect Cervical Internal Carotid Artery Dissections.

The ICA is the most common site of cervical artery dissection. Prompt and reliable identification of the mural hematoma is warranted when a dissection is clinically suspected. The purpose of this study was to assess to capacity of a standard DWI sequence acquired routinely on the brain to detect dissecting hematoma related to cervical ICA dissections. This was a retrospective study of a cohort of 110 patients younger than 55 years of age (40 women; mean age, 46.79 years) admitted at the acute phase of a neurologic deficit, headache, or neck pain and investigated by at least a standard 3T diffusion-weighted sequence of the brain. Among them were 50 patients (14 women; mean age, 46.72 years) with subsequently confirmed ICA dissection. In the whole anonymized cohort, both a senior and junior radiologist separately assessed, on the DWI sequences only, the presence of a crescent-shaped or circular hypersignal projecting on the subpetrosal segment of the ICA arteries, assuming that it would correspond to a mural hematoma related to an ICA dissection. The senior radiologist found 46 subpetrosal hyperintensities in 43/50 patients with ICA dissection and none in patients without dissection (sensitivity, 86%; specificity, 100%). The junior radiologist found 48 subpetrosal hyperintensities in 45/50 patients with dissection and none in patients without dissection (sensitivity, 90%; specificity, 100%). In our cohort, a standard DWI sequence performed on the brain at the acute phase of a stroke or for a clinical suspicion of dissection detected nearly 90% of cervical ICA dissections.

Ultrasound Computer-Aided Diagnosis (CAD) Based on the Thyroid Imaging Reporting and Data System (TI-RADS) to Distinguish Benign from Malignant Thyroid Nodules and the Diagnostic Performance of Radiologists with Different Diagnostic Experience.

BackgroundThe diagnosis of thyroid cancer and distinguishing benign from malignant thyroid nodules by junior radiologists can be challenging. This study aimed to develop a computer-aided diagnosis (CAD) system based on the Thyroid Imaging Reporting and Data System (TI-RADS) to distinguish benign from malignant thyroid nodules by analyzing ultrasound images to improve the diagnostic performance of junior radiologists.Material/MethodsA modified TI-RADS based on a convolutional neural network (CNN) was used to develop the CAD system. This retrospective study reviewed 789 thyroid nodules from 695 patients and included radiologists with different diagnostic experience. Five study groups included the CAD group, the junior radiologist group, the intermediate-level radiologist group, the senior radiologist group, and the group in which the junior radiologist used the CAD system. The ultrasound findings were reviewed and compared with the histopathology diagnosis.ResultsThe CAD system for the diagnosis of thyroid cancer showed an accuracy of 80.35%, a sensitivity of 80.64%, a specificity of 80.13%, a positive predictive value (PPV) of 76.02%, a negative predictive value (NPV) of 84.12%, and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.87. The accuracy of the junior radiologists in diagnosing thyroid cancer using CAD was similar to that of intermediate-level radiologists (79.21% vs. 77.57%; P=0.427).ConclusionsThe use of ultrasound CAD based on the TI-RADS showed potential for distinguishing between benign and malignant thyroid nodules and improved the diagnostic performance of junior radiologists.

A Novel Interpretable Computer-Aided Diagnosis System of Thyroid Nodules on Ultrasound Based on Clinical Experience

Computer-aided diagnosis systems (CADs) present valuable second opinions to radiologists in diagnosis. Many studies on thyroid nodules have proposed various CADs to provide a binary result, benignity or malignancy, for doctors, ignoring interpretability of more ultrasonic features that could be more useful. We develop an interpretable CADs (iCADS) that utilizes deep-learning networks' classification power and interpretability potential of clinical guidelines, like TIRADS, a well-established scale for thyroid nodules. iCADS incorporates a main neural-networks model and six neural-network based interpreters. The outputs of the six interpreters are compared with TIRADS guidelines and the matched result will form a report, more than a benignity or malignancy result, for radiologists. Clinical images of 16,946 thyroid nodules from 5,885 patients were used to train the proposed iCADS. An extra experimental data set containing 501 images were used to test the performance of the model. For better illustrating the assistant ability of iCADS, we also recruited ten junior radiologists to make diagnosis decisions with or without the help of different versions of iCADS. The experiments demonstrated that iCADS can largely improve junior radiologists diagnosis with the help of interpreter strategy. These experiments are also the very first attempt to evaluate the effect of interpretability of deep-learning based CADs in clinical practice. Comparison experiments with other deep-learning based CADs and traditional CADs indicated that the interpreter strategy can easily be combined to other intelligent CADs without the loss of performance. The framework of iCADS can also inspire more research on the development of CADs.

Quantification of the Asymmetry between Right and Left Cerebral Lateral Ventricles by Indexing Methods

To retrospectively analyze hydrocephalus and lateral ventricular asymmetry by measuring the midline shift, Evans index, right and left semi-Evans index, Right and left semi-callosal angles, superior ophthalmic vein (SOV) diameters, and to compare these findings with the control group. The study included 93 cases with hydrocephalus, 80 cases with asymmetrical lateral ventricles (ALV) and 83 control cases with normal findings who had cranial magnetic resonance imaging (MRI) in our institute between the years 2011 and 2016. A senior and junior radiologist analyzed the images and performed the measurements, and the results were compared. The cut-off points for the Evans index and right and left semi-Evans indexes for differentiating hydrocephalus and ALV were calculated as 28.68%, 30.77%, and 30.88%, respectively. The septum shift degree was significantly higher in the ALV group compared to the control group (P = 0.010; P < 0.05). ALV or hydrocephalus were not found to be associated with SOV enlargement. The SOVs were not found to be enlarged ipsilateral to asymmetrically enlarged lateral ventricle. Lower Evans and right and left semi-Evans indices are seen in ALV. Despite some limitations, the Evans index could still be used to differentiate hydrocephalus and ALV. Mild ALV is mostly associated with an off-midline septum. SOV diameter and enlargement are not indicators of hydrocephalus or ALV.

A deep residual learning network for predicting lung adenocarcinoma manifesting as ground-glass nodule on CT images.

To develop a deep learning-based artificial intelligence (AI) scheme for predicting the likelihood of the ground-glass nodule (GGN) detected on CT images being invasive adenocarcinoma (IA) and also compare the accuracy of this AI scheme with that of two radiologists. First, we retrospectively collected 828 histopathologically confirmed GGNs of 644 patients from two centers. Among them, 209 GGNs are confirmed IA and 619 are non-IA, including 409 adenocarcinomas in situ and 210 minimally invasive adenocarcinomas. Second, we applied a series of pre-preprocessing techniques, such as image resampling, rescaling and cropping, and data augmentation, to process original CT images and generate new training and testing images. Third, we built an AI scheme based on a deep convolutional neural network by using a residual learning architecture and batch normalization technique. Finally, we conducted an observer study and compared the prediction performance of the AI scheme with that of two radiologists using an independent dataset with 102 GGNs. The new AI scheme yielded an area under the receiver operating characteristic curve (AUC) of 0.92 ± 0.03 in classifying between IA and non-IA GGNs, which is equivalent to the senior radiologist's performance (AUC 0.92 ± 0.03) and higher than the score of the junior radiologist (AUC 0.90 ± 0.03). The Kappa value of two sets of subjective prediction scores generated by two radiologists is 0.6. The study result demonstrates using an AI scheme to improve the performance in predicting IA, which can help improve the development of a more effective personalized cancer treatment paradigm. • The feasibility of using a deep learning method to predict the likelihood of the ground-glass nodule being invasive adenocarcinoma. • Residual learning-based CNN model improves the performance in classifying between IA and non-IA nodules. • Artificial intelligence (AI) scheme yields higher performance than radiologists in predicting invasive adenocarcinoma.

An inception module CNN classifiers fusion method on pulmonary nodule diagnosis by signs

A “sign” on a lung CT image refers to a radiologic finding that suggests a pathological progression of some specific disease. Analysis of CT signs is helpful to understand the pathological origin of the lesion. In-depth study of lung nodules classification with different CT signs will help to distinguish benign and malignant nodules more clearly and accurately. To this end, we propose an Inception module-based ensemble classification method for pulmonary nodule diagnosis with different nodule signs. We first construct a Convolutional Neural Network (CNN) classifier adopting Inception modules and pre-train it on ImageNet. We then fine-tune this pre-trained classifier on 10 different lung nodule sign sample sets, and fuse these 10 classifiers with an artificial immune ensemble algorithm. The overall sensitivity, specificity, and accuracy of our proposed Artificial Immune Algorithm-based Inception Networks Fusion (AIA-INF) algorithm are 82.22%, 93.17%, and 88.67%, respectively, which are significantly higher than those of the alternative Bagging and Boosting methods. The experimental results show that our Inception-based ensemble classifier offers promising performance, and compared with other CADx systems, this scheme can offer a more detailed reference for diagnosis, and can be valuable for junior radiologist training.

Learning curve of liver stiffness measurement using a new hybrid machine composed of transient elastography interfaced with ultrasound.

To assess the learning curve for performing reliable liver stiffness measurements using a new hybrid machine composed of transient elastography (TE) interfaced with an ultrasound device for radiographers and radiologists with different levels of expertise in ultrasound imaging. Ten novice operators who had never performed TE measurements were prospectively evaluated from April to October 2018: senior radiologists, young radiologists, fellows, radiographers, and residents, with different levels of experience in abdominal ultrasound imaging. All operators had a short theoretical training followed by a training session under supervision in three patients. Then, each operator had to perform TE in 50 consecutive patients with chronic liver disease, using beforehand ultrasound examination to select measurement area in the right liver lobe, and if needed, the XL probe. Percentages of failures and reliable measurements were compared. The rates of failures of measurements, poorly reliable, reliable, and very reliable results, were of 4.2% (21/500), 2.4% (12/500), 47.6% (238/500), and 45.8% (229/500), respectively. The rates of reliable plus very reliable results were excellent, ranging from 91 to 96% among all the subgroups. The rates of very reliable, reliable, and unreliable results did not differ between operator subgroups and especially between junior radiologists, senior radiologists, and radiographers. No breaking point was observed in the interquartile range/median values over time. TE interfaced with ultrasound in this hybrid machine presents no learning curve effect. After a short initial training session, a novice observer is able to perform high rates of reliable and very reliable TE measurements. • When performing liver stiffness measurements using a new hybrid machine composed of transient elastography interfaced with ultrasound, the rate of failures of measurements is very low, below 5%. • After a short training session and using ultrasound planning, a novice operator, whatever its expertise in ultrasound imaging, is capable of performing high rates of reliable and very reliable measurements. • No learning curve is needed for performing reliable liver stiffness measurements using this new hybrid machine.

Improving diagnosis accuracy of brain volume abnormalities during childhood with an automated MP2RAGE-based MRI brain segmentation

Background and purposeIt can be challenging to depict brain volume abnormalities in the pediatric population on magnetic resonance imaging (MRI). The aim of the study was to evaluate the inter-radiologist reliability in brain MRI interpretation, including brain volume assessment and the efficiency of an automated brain segmentation. Materials and methodsWe performed a single-center prospective study including 44 patients aged six months to five years recruited from the University Hospital, having a 1.5T brain MRI using a MP2RAGE sequence. All MRI were randomly and blindly reviewed by one junior and two senior pediatric radiologists. Inter-observer agreements were assessed using Fleiss’ kappa coefficient. Brain volumetry (total intracranial volume (TIV), brain parenchyma, and cerebrospinal fluid volumes) was estimated using the MorphoBox prototype. Clinical head circumference (HC) and z scores were reported. A Pearson correlation coefficient was calculated between brain volumes with HC. ResultsTwenty-four brain MRI examinations were normal and twenty were pathological. Brain volume abnormalities were poorly detected by junior and senior radiologists: sensitivities 16.67% [confidence interval 4.7–44.8], 33.33% [13–60] and 30.7% [12–58] and specificities 93.75% [79–98], 84.38% [68–93] and 77% [60–88], respectively. Brain volume apart, interobserver kappa coefficients were 0.93 between junior and seniors as well as between seniors. Brain volumes were significantly correlated with HC (P<0.0001). In patients with normal MRI, brain parenchyma volumes increased regularly with age. Low brain volume was easier to identify with automated quantification. ConclusionBrain volume was poorly appreciated by radiologists. The fully automated brain segmentation used can provide quantitative data to better diagnose, describe, and follow-up brain volume abnormalities.

3D Automated Breast Ultrasound System: Comparison of Interpretation Time of Senior Versus Junior Radiologist.

This study aimed to compare the automated breast ultrasound system (ABUS) reading time of breast radiologist to a radiology resident independent of the clinical outcomes. One hundred women who underwent screening ABUS between July and August 2017 were reviewed retrospectively. Each study was examined sequentially by a breast radiologist who has more than 20 years of experience in breast radiology and third year resident who has 6 months of experience in breast radiology. Data were analyzed with Spearman' correlation, Wilcoxon Signed Ranks Test and Kruskal-Wallis Test and was recorded. The mean age of patients was 42.02±11.423 years (age range16-66). The average time for senior radiologist was 223.36±84.334 seconds (min 118 max 500 seconds). The average time for junior radiologist was 269.48±82.895 seconds (min 150 max 628 seconds). There was a significant difference between the mean time of two radiologists (p=0.00001). There was a significant difference regarding the decrease in the reading time throughout study with the increase of number of cases read by the breast radiologist (p<0.05); but not with the resident radiologist (p=0.687). There was a correlation between BI-RADS category and reading time for both the breast radiologist and the resident (p=0.002, p=0.00043 respectively) indicating that patients who had findings caused longer reading times. ABUS reading time may differ according to the experience of the user, however the times of an experienced and non-experienced user is comparable.

Interobserver agreement of whole-body magnetic resonance imaging is superior to whole-body computed tomography for assessing disease burden in patients with multiple myeloma

ObjectivesWhole-body MRI (WB-MRI) is recommended by the International Myeloma Working Group for all patients with asymptomatic myeloma and solitary plasmacytoma and by the UK NICE guidance for all patients with suspected myeloma. Some centres unable to offer WB-MRI offer low-dose whole-body CT (WB-CT). There are no studies comparing interobserver agreement and disease detection of contemporary WB-MRI (anatomical imaging and DWI) versus WB-CT. Our primary aim is to compare the interobserver agreement between WB-CT and WB-MRI in the diagnosis of myeloma.MethodsConsecutive patients with newly diagnosed myeloma imaged with WB-MRI and WB-CT were prospectively reviewed. For each body region and modality, two experienced and two junior radiologists scored disease burden with final scores by consensus. Intraclass correlation coefficients (ICC), median scores, Wilcoxon signed-rank test and Spearman’s correlation coefficients were calculated.ResultsThere was no significant difference in overall observer scores between WB-MRI and WB-CT (p = 0.87). For experienced observers, interobserver agreement for WB-MRI was superior to WB-CT overall and for each region, without overlap in whole-skeleton confidence intervals (ICC 0.98 versus 0.77, 95%CI 0.96–0.99 versus 0.45–0.91). For inexperienced observers, although there is a trend for a better interobserver score for the whole skeleton on WB-MRI (ICC 0.95, 95%CI 0.72–0.98) than on WB-CT (ICC 0.72, 95%CI 0.34–0.88), the confidence intervals overlap.ConclusionsWB-MRI offers excellent interobserver agreement which is superior to WB-CT for experienced observers. Although the overall burden was similar across both modalities, patients with lower disease burdens where MRI could be advantageous are not included in this series.Key Points• Whole-body MRI is recommended by the International Myeloma Working Group for patients with multiple myeloma and solitary plasmacytoma and by the NICE guidance for those with suspected multiple myeloma.• Some centres unable to offer whole-body MRI (WB-MRI) offer low-dose whole-body CT (WB-CT).• This prospective study demonstrates that contemporary WB-MRI (with anatomical sequences and DWI) provides better interobserver agreement in assessing myeloma disease burden for the whole skeleton and across any individual body region in myeloma patients when compared with low-dose whole-body CT.

Interpretation of Basic Clinical Images: How Are Surgical Residents Performing Compared to Other Trainees?

During medical training students, residents, and fellows learn how to accurately interpret basic radiographic images. This skill is mostly utilized by physicians in the acute and critical care settings. It is unclear whether surgical residents' interpretation skills differ from that of other trainees. A 30-question online quiz was developed to evaluate trainees' skills in interpreting images using various radiologic modalities. The participating cohort included (1) medical students (MS), (2) general surgery residents (GST), internal medicine residents and fellows (IMT), and radiology trainees (RT). The impact of residency specialty and level of training on performance was evaluated. A total of 69 postgraduate trainees and 19 MS enrolled in the online quiz. The average score was 67.6% (±16.6). GST scored higher than IMT (74.2% ± 10.7% vs. 67.9% ± 11.3%, p = 0.038); however, they were equally proficient to RT. MS had the lowest interpretation accuracy rates compared to postgraduate trainees (57.4% ± 16.8%, p < 0.001). On different radiographic modalities, junior GST performance was comparable to MS, JR-IMT, and Junior Radiology Trainees (JR-RT). On computed tomography (CT) body, GST (83.1% ± 15.7%) scored higher than IMT (70.3% ± 17.7%, p = 0.026) and MS (61.7% ± 23.4%, p < 0.001). Similar findings were demonstrated on ultrasound modality. A difference in performance was not evident for X-rays, CT head, and tubes/lines localization images. GST were able to correctly interpret 74.2% of basic clinical images. Although superior in the evaluation of pathologies seen on CT body and ultrasound, GST have comparable performance to other trainees in X-rays, tube/line localization images, and CT head. Integration of radiology education in surgical training may enhance performance and potentially improve patient care.

Evaluating a Fully Automated Pulmonary Nodule Detection Approach and Its Impact on Radiologist Performance.

To compare sensitivity in the detection of lung nodules between the deep learning (DL) model and radiologists using various patient population and scanning parameters and to assess whether the radiologists' detection performance could be enhanced when using the DL model for assistance. A total of 12 754 thin-section chest CT scans from January 2012 to June 2017 were retrospectively collected for DL model training, validation, and testing. Pulmonary nodules from these scans were categorized into four types: solid, subsolid, calcified, and pleural. The testing dataset was divided into three cohorts based on radiation dose, patient age, and CT manufacturer. Detection performance of the DL model was analyzed by using a free-response receiver operating characteristic curve. Sensitivities of the DL model and radiologists were compared by using exploratory data analysis. False-positive detection rates of the DL model were compared within each cohort. Detection performance of the same radiologist with and without the DL model were compared by using nodule-level sensitivity and patient-level localization receiver operating characteristic curves. The DL model showed elevated overall sensitivity compared with manual review of pulmonary nodules. No significant dependence regarding radiation dose, patient age range, or CT manufacturer was observed. The sensitivity of the junior radiologist was significantly dependent on patient age. When radiologists used the DL model for assistance, their performance improved and reading time was reduced. DL shows promise to enhance the identification of pulmonary nodules and benefit nodule management.© RSNA, 2019Supplemental material is available for this article.

Precise diagnosis of intracranial hemorrhage and subtypes using a three-dimensional joint convolutional and recurrent neural network

ObjectivesTo evaluate the performance of a novel three-dimensional (3D) joint convolutional and recurrent neural network (CNN-RNN) for the detection of intracranial hemorrhage (ICH) and its five subtypes (cerebral parenchymal, intraventricular, subdural, epidural, and subarachnoid) in non-contrast head CT.MethodsA total of 2836 subjects (ICH/normal, 1836/1000) from three institutions were included in this ethically approved retrospective study, with a total of 76,621 slices from non-contrast head CT scans. ICH and its five subtypes were annotated by three independent experienced radiologists, with majority voting as reference standard for both the subject level and the slice level. Ninety percent of data was used for training and validation, and the rest 10% for final evaluation. A joint CNN-RNN classification framework was proposed, with the flexibility to train when subject-level or slice-level labels are available. The predictions were compared with the interpretations from three junior radiology trainees and an additional senior radiologist.ResultsIt took our algorithm less than 30 s on average to process a 3D CT scan. For the two-type classification task (predicting bleeding or not), our algorithm achieved excellent values (≥ 0.98) across all reporting metrics on the subject level. For the five-type classification task (predicting five subtypes), our algorithm achieved > 0.8 AUC across all subtypes. The performance of our algorithm was generally superior to the average performance of the junior radiology trainees for both two-type and five-type classification tasks.ConclusionsThe proposed method was able to accurately detect ICH and its subtypes with fast speed, suggesting its potential for assisting radiologists and physicians in their clinical diagnosis workflow.Key Points• A 3D joint CNN-RNN deep learning framework was developed for ICH detection and subtype classification, which has the flexibility to train with either subject-level labels or slice-level labels.• This deep learning framework is fast and accurate at detecting ICH and its subtypes.• The performance of the automated algorithm was superior to the average performance of three junior radiology trainees in this work, suggesting its potential to reduce initial misinterpretations.

Confessions of a Female Junior Radiology Faculty Member

Confessions of a Female Junior Radiology Faculty Member

678 - Reproducibility and interobserver variability of the RENAL nephrometry score: Senior versus junior radiologists

678 - Reproducibility and interobserver variability of the RENAL nephrometry score: Senior versus junior radiologists