Radiologist Assessment Research Articles

Image Recovery from Synthetic Noise Artifacts in CT Scans Using Modified U-Net.

Computed Tomography (CT) is commonly used for cancer screening as it utilizes low radiation for the scan. One problem with low-dose scans is the noise artifacts associated with low photon count that can lead to a reduced success rate of cancer detection during radiologist assessment. The noise had to be removed to restore detail clarity. We propose a noise removal method using a new model Convolutional Neural Network (CNN). Even though the network training time is long, the result is better than other CNN models in quality score and visual observation. The proposed CNN model uses a stacked modified U-Net with a specific number of feature maps per layer to improve the image quality, observable on an average PSNR quality score improvement out of 174 images. The next best model has 0.54 points lower in the average score. The score difference is less than 1 point, but the image result is closer to the full-dose scan image. We used separate testing data to clarify that the model can handle different noise densities. Besides comparing the CNN configuration, we discuss the denoising quality of CNN compared to classical denoising in which the noise characteristics affect quality.

Radiomics Signature Using Manual Versus Automated Segmentation for Lymph Node Staging of Bladder Cancer

BackgroundBladder cancer (BC) treatment algorithms depend on accurate tumor staging. To date, computed tomography (CT) is recommended for assessment of lymph node (LN) metastatic spread in muscle-invasive and high-risk BC. However, the diagnostic efficacy of radiologist-evaluated CT imaging studies is limited. ObjectiveTo evaluate the performance of quantitative radiomics signatures for detection of LN metastases in BC. Design, setting, and participantsOf 1354 patients with BC who underwent radical cystectomy (RC) with lymphadenectomy who were screened, 391 with pathological nodal staging (pN0: n = 297; pN+: n = 94) were included and randomized into training (n = 274) and test (n = 117) cohorts. Pelvic LNs were segmented manually and automatically. A total of 1004 radiomics features were extracted from each LN and a machine learning model was trained to assess pN status using histopathology labels as the ground truth. Outcome measurements and statistical analysisRadiologist assessment was compared to radiomics-based analysis using manual and automated LN segmentations for detection of LN metastases in BC. Statistical analysis was performed using the receiver operating characteristics curve method and evaluated in terms of sensitivity, specificity, and area under the curve (AUC). Results and limitationsIn total, 1845 LNs were manually segmented. Automated segmentation correctly located 361/557 LNs in the test cohort. Manual and automatic masks achieved an AUC of 0.80 (95% confidence interval [CI] 0.69–0.91; p = 0.64) and 0.70 (95% CI: 0.58–0.82; p = 0.17), respectively, in the test cohort compared to radiologist assessment, with an AUC of 0.78 (95% CI 0.67–0.89). A combined model of a manually segmented radiomics signature and radiologist assessment reached an AUC of 0.81 (95% CI 0.71–0.92; p = 0.63). ConclusionsA radiomics signature allowed discrimination of nodal status with high diagnostic accuracy. The model based on manual LN segmentation outperformed the fully automated approach. Patient summaryFor patients with bladder cancer, evaluation of computed tomography (CT) scans before surgery using a computer-based method for image analysis, called radiomics, may help in standardizing and improving the accuracy of assessment of lymph nodes. This could be a valuable tool for optimizing treatment options.

Differentiation between spinal multiple myeloma and metastases originated from lung using multi-view attention-guided network.

PurposeMultiple myeloma (MM) and metastasis originated are the two common malignancy diseases in the spine. They usually show similar imaging patterns and are highly demanded to differentiate for precision diagnosis and treatment planning. The objective of this study is therefore to construct a novel deep-learning-based method for effective differentiation of two diseases, with the comparative study of traditional radiomics analysis.MethodsWe retrospectively enrolled a total of 217 patients with 269 lesions, who were diagnosed with spinal MM (79 cases, 81 lesions) or spinal metastases originated from lung cancer (138 cases, 188 lesions) confirmed by postoperative pathology. Magnetic resonance imaging (MRI) sequences of all patients were collected and reviewed. A novel deep learning model of the Multi-view Attention-Guided Network (MAGN) was constructed based on contrast-enhanced T1WI (CET1) sequences. The constructed model extracts features from three views (sagittal, coronal and axial) and fused them for a more comprehensive differentiation analysis, and the attention guidance strategy is adopted for improving the classification performance, and increasing the interpretability of the method. The diagnostic efficiency among MAGN, radiomics model and the radiologist assessment were compared by the area under the receiver operating characteristic curve (AUC).ResultsAblation studies were conducted to demonstrate the validity of multi-view fusion and attention guidance strategies: It has shown that the diagnostic model using multi-view fusion achieved higher diagnostic performance [ACC (0.79), AUC (0.77) and F1-score (0.67)] than those using single-view (sagittal, axial and coronal) images. Besides, MAGN incorporating attention guidance strategy further boosted performance as the ACC, AUC and F1-scores reached 0.81, 0.78 and 0.71, respectively. In addition, the MAGN outperforms the radiomics methods and radiologist assessment. The highest ACC, AUC and F1-score for the latter two methods were 0.71, 0.76 & 0.54, and 0.69, 0.71, & 0.65, respectively.ConclusionsThe proposed MAGN can achieve satisfactory performance in differentiating spinal MM between metastases originating from lung cancer, which also outperforms the radiomics method and radiologist assessment.

Artificial intelligence software in pulmonary nodule assessment.

This study tests the impact of the addition of autonomous computed tomography (CT) interpreting software to radiologist assessment of pulmonary nodules. Computed tomography scans for nodule assessment were identified retrospectively. Lung cancer risk factors, initial radiologist (RAD) report, Philips Lung Nodule software report (computer-aided nodule (CAD)) and radiologist report following the review of CT images and CAD (RAD + CAD) were collected. Follow-up recommendations based on current guidelines were derived from each report. In all, 100 patients were studied. Median maximal diameter of the largest nodule reported by RAD and RAD + CAD were similar at 10.0 and 9.0 mm, respectively (p = 0.06) but were reported as larger by CAD at 11.8 mm (p < 0.001). Follow-up recommendations derived from RAD + CAD were less intensive in 23 (23%) and more intensive in 34 (34%) than that of RAD. This study suggests that autonomous software use can alter radiologist assessment of pulmonary nodules such that suggested follow-up is altered.

Deep-learning-based 3D super-resolution MRI radiomics model: superior predictive performance in preoperative T-staging of rectal cancer

ObjectivesTo investigate the feasibility and efficacy of a deep-learning (DL)-based three-dimensional (3D) super-resolution (SR) MRI radiomics model for preoperative T-staging prediction in rectal cancer (RC).MethodsSeven hundred six eligible RC patients (T1/2 = 287, T3/4 = 419) were retrospectively enrolled in this study and chronologically allocated into a training cohort (n = 565) and a validation cohort (n = 141). We conducted a deep-transfer-learning network on high-resolution (HR) T2-weighted imaging (T2WI) to enhance the z-resolution of the images and acquired the preoperative SRT2WI. The radiomics models named modelHRT2 and modelSRT2 were respectively constructed with high-dimensional quantitative features extracted from manually segmented volume of interests of HRT2WI and SRT2WI through the Least Absolute Shrinkage and Selection Operator method. The performances of the models were evaluated by ROC, calibration, and decision curves.ResultsModelSRT2 outperformed modelHRT2 (AUC 0.869, sensitivity 71.1%, specificity 93.1%, and accuracy 83.3% vs. AUC 0.810, sensitivity 89.5%, specificity 70.1%, and accuracy 77.3%) in distinguishing T1/2 and T3/4 RC with significant difference (p < 0.05). Both radiomics models achieved higher AUCs than the expert radiologists (0.685, 95% confidence interval 0.595–0.775, p < 0.05). The calibration curves confirmed high goodness of fit, and the decision curve analysis revealed the clinical value.ConclusionsModelSRT2 yielded superior predictive performance in preoperative RC T-staging by comparison with modelHRT2 and expert radiologists’ visual assessments.Key Points• For the first time, DL-based 3D SR images were applied in radiomics analysis for clinical utility.• Compared with the visual assessment of expert radiologists and the conventional radiomics model based on HRT2WI, the SR radiomics model showed a more favorable capability in helping clinicians assess the invasion depth of RC preoperatively.• This is the largest radiomics study for T-staging prediction in RC.

Artificial intelligence to complement rather than replace radiologists in breast screening

Artificial intelligence to complement rather than replace radiologists in breast screening

Evaluation of the Combination of Artificial Intelligence and Radiologist Assessments to Interpret Malignant Architectural Distortion on Mammography.

PurposeTo compare the mammographic malignant architectural distortion (AD) detection performance of radiologists who read mammographic examinations unaided versus those who read these examinations with the support of artificial intelligence (AI) systems.Material and MethodsThis retrospective case-control study was based on a double-reading of clinical mammograms between January 2011 and December 2016 at a large tertiary academic medical center. The study included 177 malignant and 90 benign architectural distortion (AD) patients. The model was built based on the ResNeXt-50 network. Algorithms used deep learning convolutional neural networks, feature classifiers, image analysis algorithms to depict AD and output a score that translated to malignant. The accuracy for malignant AD detection was evaluated using area under the curve (AUC).ResultsThe overall AUC was 0.733 (95% CI, 0.673-0.792) for Reader First-1, 0.652 (95% CI, 0.586-0.717) for Reader First-2, and 0.655 (95% CI, 0.590-0.719) for Reader First-3. and the overall AUCs for Reader Second-1, 2, 3 were 0.875 (95% CI, 0.830-0.919), 0.882 (95% CI, 0.839-0.926), 0.884 (95% CI, 0.841-0.927),respectively. The AUCs for all the reader-second radiologists were significantly higher than those for all the reader-first radiologists (Reader First-1 vs. Reader Second-1, P= 0.004). The overall AUC was 0.792 (95% CI, 0.660-0.925) for AI algorithms. The combination assessment of AI algorithms and Reader First-1 achieved an AUC of 0.880 (95% CI, 0.793-0.968), increased than the Reader First-1 alone and AI algorithms alone. AI algorithms alone achieved a specificity of 61.1% and a sensitivity of 80.6%. The specificity for Reader First-1 was 55.5%, and the sensitivity was 86.1%. The results of the combined assessment of AI and Reader First-1 showed a specificity of 72.7% and sensitivity of 91.7%. The performance showed significant improvements compared with AI alone (p<0.001) as well as the reader first-1 alone (p=0.006).ConclusionWhile the single AI algorithm did not outperform radiologists, an ensemble of AI algorithms combined with junior radiologist assessments were found to improve the overall accuracy. This study underscores the potential of using machine learning methods to enhance mammography interpretation, especially in remote areas and primary hospitals.

Optimizing the radiomics-machine-learning model based on non-contrast enhanced CT for the simplified risk categorization of thymic epithelial tumors: A large cohort retrospective study

PurposeThis study aimed to establish and compare the radiomics machine learning (ML) models based on non-contrast enhanced computed tomography (NECT) and clinical features for predicting the simplified risk categorization of thymic epithelial tumors (TETs). Experimental designA total of 509 patients with pathologically confirmed TETs from January 2009 to May 2018 were retrospectively enrolled, consisting of 238 low-risk thymoma (LRT), 232 high-risk thymoma (HRT), and 39 thymic carcinoma (TC), and were divided into training (n = 433) and testing cohorts (n = 76) according to the admission time. Volumes of interest (VOIs) covering the whole tumor were manually segmented on preoperative NECT images. A total of 1218 radiomic features were extracted from the VOIs, and 4 clinical variables were collected from the hospital database. Fourteen ML models, along with varied feature selection strategies, were used to establish triple-classification models using the radiomic features (radiomic models), while clinical-radiomic models were built after combining with the clinical variables. The diagnostic accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) of radiologist assessment, the radiomic and clinical-radiomic models were evaluated on the testing cohort. ResultsThe Support Vector Machine (SVM) clinical-radiomic model demonstrated the highest AUC of 0.841 (95% CI 0.820 to 0.861) on the cross-validation result and reached an AUC of 0.844 (95% CI 0.793 to 0.894) in the testing cohort. For the one-vs-rest question of LRT vs HRT + TC, the sensitivity, specificity, and accuracy reached 80.00%, 63.41%, and 71.05%, respectively. For HRT vs LRT + TC, they reached 60.53%, 78.95%, and 69.74%. For TC vs LRT + HRT they reached 33.33%, 98.63%, and 96.05%, respectively. Compared with the radiomic models, superior diagnostic efficacy was demonstrated for most clinical-radiomics models, and the AUC of the Bernoulli Naive Bayes model was significantly improved. Radiologist2′s assessment achieved a higher AUC of 0.813 (95% CI: 0.756–0.8761) than other radiologists, which was slightly lower than the SVM clinical-radiomic model. Combined with other evaluation indicators, SVM, as the best ML model, demonstrated the potential of predicting the simplified risk categorization of TETs with superior predictive performance to that of radiologists’ assessment. ConclusionMost of the ML models are promising in predicting the simplified TETs risk categorization with superior efficacy to that of radiologists’ assessment, especially the SVM models, demonstrated the integration of ML with NECT may be valuable in aiding the diagnosis and treatment planning.

Diagnosing Ovarian Cancer on MRI: A Preliminary Study Comparing Deep Learning and Radiologist Assessments.

Simple SummaryAs a preliminary experiment to explore the possibility of clinical application as a future reading assist, we present CNNs for the diagnosis of ovarian carcinomas and borderline tumors on MRI, including T2WI, DWI, ADC map, and CE-T1WI, and compare their diagnostic performance with interpretations by experienced radiologists. CNNs were trained using 1798 images from 146 patients and 1865 images from 219 patients with malignant tumors, including borderline tumors, and non-malignant lesions, respectively, for each MRI sequence and tested with 48 and 52 images of patients with malignant and non-malignant lesions. The CNN of each sequence had a sensitivity of 0.77–0.85, specificity of 0.77–0.92, accuracy of 0.81–0.87, and an AUC of 0.83–0.89, demonstrating diagnostic performances that were non-inferior to those of experienced radiologists, and the CNN showed the highest diagnostic performance on the ADC map for each sequence (specificity = 0.85; sensitivity = 0.77; accuracy = 0.81; AUC = 0.89).Background: This study aimed to compare deep learning with radiologists’ assessments for diagnosing ovarian carcinoma using MRI. Methods: This retrospective study included 194 patients with pathologically confirmed ovarian carcinomas or borderline tumors and 271 patients with non-malignant lesions who underwent MRI between January 2015 and December 2020. T2WI, DWI, ADC map, and fat-saturated contrast-enhanced T1WI were used for the analysis. A deep learning model based on a convolutional neural network (CNN) was trained using 1798 images from 146 patients with malignant tumors and 1865 images from 219 patients with non-malignant lesions for each sequence, and we tested with 48 and 52 images of patients with malignant and non-malignant lesions, respectively. The sensitivity, specificity, accuracy, and AUC were compared between the CNN and interpretations of three experienced radiologists. Results: The CNN of each sequence had a sensitivity of 0.77–0.85, specificity of 0.77–0.92, accuracy of 0.81–0.87, and an AUC of 0.83–0.89, and it achieved a diagnostic performance equivalent to the radiologists. The CNN showed the highest diagnostic performance on the ADC map among all sequences (specificity = 0.85; sensitivity = 0.77; accuracy = 0.81; AUC = 0.89). Conclusion: The CNNs provided a diagnostic performance that was non-inferior to the radiologists for diagnosing ovarian carcinomas on MRI.

DIAGNOSTIC PERFORMANCE OF DEEP LEARNING ARCHITECTURES IN THE DIAGNOSIS OF SJOGRENS SYNDROME- A SYSTEMATIC LITERTAURE REVIEW

Background: Sjogren's Syndrome (SS) is a chronic, multifactorial autoimmune disease, characterized by clinical symptoms of dry mouth and dry eyes, due to chronic lymphocytic destruction of salivary and lacrimal glands, respectively. Proper diagnosis is a key towards better outcome. Recently introduced deep learning systems have ability to reflect the complexity of condition, with an aim to bring personalized medicine closer to the patients. Aim: The aim of this systematic review is to compile evidence-based studies pertaining to diagnostic performance of DL system and its algorithms in diagnosis of monitoring of SS. Materials and method: Computerized literature search was performed to select eligible articles from the following databases: PUBMED [MEDLINE], SCOPUS, SCIENCE DIRECT and COCHRANE DATABASE using specific keywords. The search was limited to articles published as full text in English, which were screened by two authors for eligibility. Results: Four studies satisfied our inclusion criteria, that suggested it to have high diagnostic accuracy when compared to inexperienced radiologist, but equivalent to those of experienced radiologists. Two studies found accuracy, sensitivity, and specificity of DL systems to be 89.5%, 90.0%, and 89.0%, for USG salivary gland images respectively whereas for CT images, the accuracy, sensitivity, and specificity was observed to be 96.0%,100% and 92.0%, respectively, and the diagnostic performance was higher from an inexperienced radiologist (p &lt; 0.0001). Conclusion: DL systems have the potential to provide useful diagnostic support to inexperienced radiologists in assessment of images for the presence of characteristic features of SS. They could assist the radiologists in automated segmentation of salivary glands, and enables feature extraction in a reduced time with reduced risk of cognitive errors.

Abstract PD4-01: Response according to revised RANO criteria is associated with overall survival in breast cancer patients with leptomeningeal metastasis

Abstract Introduction: Leptomeningeal metastasis (LM) is a severe complication of many solid tumors generally associated with dismal prognosis. Limited therapeutic options are available in this setting and assessment of response to treatment currently lacks validated tools. In 2017, the Response Assessment in Neuro-Oncology (RANO) LM Working Group proposed a standardized scorecard to aid in the evaluation of MRI findings in LM, which was revised and simplified in a subsequent 2019 version. To date, the prognostic impact of response to treatment evaluated according to these criteria has not been assessed in breast cancer (BC). In this study, we aim to validate this tool in a multicentric cohort of BC patients (pts) diagnosed with LM.Methods: Pts with LM from BC diagnosed at Istituto Oncologico Veneto (Padova, Italy) and Montpellier Regional Cancer Institute (Montpellier, France) between December 2005 and June 2018 were identified. Clinicopathological characteristics were collected from medical charts. Baseline and follow-up MRI scans were centrally reviewed by a dedicated radiologist and response assessment was evaluated according to the 2019 revised RANO LM response criteria (Le Rhun E. et al., Neuro-Oncology 2019).Results: A total of 225 pts diagnosed with BC-related LM were identified; among these, 142 had an available baseline brain MRI imaging and were included in the present study. Among these 142 pts, 78 had a positive cerebrospinal fluid (CSF) cytology and were classified as type I according to EANO-ESMO classification (41 type IA, 9 type IB, 28 type IC); for 64 pts CSF cytology was negative (N=25) or not performed (N=39), and were classified as type II (28 type IIA, 4 type IIB, 32 type IIC). At the time of last follow-up, 138 out of 142 pts had died and median overall survival (OS) from baseline MRI was 5.0 months (95% CI 3.4-6.7). Presence/absence of positive CSF cytology did not impact OS (median OS 5.3 versus 5.2 months, log-rank p=0.406). A total of 60 pts (positive/negative/not done CSF cytology: N=36/9/15) had an evaluable follow-up MRI imaging, with drop-off mainly due to deteriorating performance status or death. In this selected subgroup of pts, a longer OS was observed (median OS 15.2 months; 95% CI 9.5-21.0). Treatment administered for BC-related LM was heterogeneous and several pts received more than one treatment modality: systemic chemotherapy (N=44), endocrine therapy (N=27), HER2-targeted treatment (N=16), intrathecal chemotherapy (N=37) and radiotherapy (N=24). Response to treatment at time of first disease re-evaluation by brain MRI imaging was as follows: complete response for 2 pts, partial response for 12 pts, stable disease for 33 pts, and progression of disease for 13 pts. Response according to 2019 RANO criteria was significantly associated with OS (p=0.029, Table 1). Conclusions: Response assessment in pts with LM is challenging. In our study, the 2019 RANO revised response score was associated with OS in patients with LM from BC, thus validating these criteria in a real-world cohort. Response according to 2019 RANO LM criteriaMedian OS (months)OS 95% CI (months)Univariate Cox Hazard RatioUnivariate Cox Hazard Ratio 95% CIp-valueComplete Response31.1NE-NE0.100.01-0.780.028Partial Response16.10.0-37.40.410.17-0.970.043Stable Disease17.98.3-27.50.450.22-0.910.026Progression of Disease9.55.9-13.1Ref.Ref.Ref. Citation Format: Gaia Griguolo, Vittoria Aldegheri, Michele Bottosso, Francesca Caumo, Stéphane Pouderoux, Federica Miglietta, William Jacot, Maria Vittoria Dieci, Amélie Darlix, Valentina Guarneri. Response according to revised RANO criteria is associated with overall survival in breast cancer patients with leptomeningeal metastasis [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr PD4-01.

Implementing Real-World RECIST-based Tumor Response Assessment in Patients With Metastatic Non-small Cell Lung Cancer

BackgroundTo accelerate drug approvals while maintaining scientific rigor in the evaluation of a therapeutic's efficacy and safety, the United States Food and Drug Administration now considers real-world data (RWD) to support New Drug Applications and expanded indications. Response Evaluation Criteria in Solid Tumors (RECIST) are the gold standard in clinical trials, but the derivation of RECIST-based treatment response from RWD is unproven. This study investigated the feasibility of implementing RECIST criteria in RWD by comparing lung cancer response assessments from RECIST-based measurement of lesions on archived radiologic films with results from medical oncologist and radiologist narratives recorded in electronic health records (EHR). Materials and MethodsResponse to index treatment via different assessment approaches was compared among 30 metastatic non-small cell lung cancer (mNSCLC) patients receiving systemic treatment (index) after progression on a platinum or anti-PD(L)-1-containing regimen. Specifically, responses based on assessments documented in the medical oncologists’ narratives were compared to a radiologist's assessments of archived images using RECIST v1.1 criteria. Each patient's best overall response was characterized as complete or partial (CR/PR), stable disease (SD), progressive disease (PD), or not evaluable (NE). ResultsSimilar distributions of best overall response and substantial concordance (77%) between medical oncologist-reported and radiologist re-assessed responses were observed. There were no instances of CR/PR to PD or PD to CR/PR discordance. ConclusionsResults suggest that accurate treatment responses, similar to RECIST, may be derived using RWD. Further validation and improvement of real-world response assessment are needed to develop a scalable real-world approach for response assessment.

An MRI-Based Radiomics Model for Predicting the Benignity and Malignancy of BI-RADS 4 Breast Lesions.

ObjectivesThe probability of Breast Imaging Reporting and Data Systems (BI-RADS) 4 lesions being malignant is 2%–95%, which shows the difficulty to make a diagnosis. Radiomics models based on magnetic resonance imaging (MRI) can replace clinicopathological diagnosis with high performance. In the present study, we developed and tested a radiomics model based on MRI images that can predict the malignancy of BI-RADS 4 breast lesions.MethodsWe retrospective enrolled a total of 216 BI-RADS 4 patients MRI and clinical information. We extracted 3,474 radiomics features from dynamic contrast-enhanced (DCE), T2-weighted images (T2WI), and diffusion-weighted imaging (DWI) MRI images. Least absolute shrinkage and selection operator (LASSO) and logistic regression were used to select features and build radiomics models based on different sequence combinations. We built eight radiomics models which were based on DCE, DWI, T2WI, DCE+DWI, DCE+T2WI, DWI+T2WI, and DCE+DWI+T2WI and a clinical predictive model built based on the visual assessment of radiologists. A nomogram was constructed with the best radiomics signature combined with patient characteristics. The calibration curves for the radiomics signature and nomogram were conducted, combined with the Hosmer-Lemeshow test.ResultsPearson’s correlation was used to eliminate 3,329 irrelevant features, and then LASSO and logistic regression were used to screen the remaining feature coefficients for each model we built. Finally, 12 related features were obtained in the model which had the best performance. These 12 features were used to build a radiomics model in combination with the actual clinical diagnosis of benign or malignant lesion labels we have obtained. The best model built by 12 features from the 3 sequences has an AUC value of 0.939 (95% CI, 0.884-0.994) and an accuracy of 0.931 in the testing cohort. The sensitivity, specificity, precision and Matthews correlation coefficient (MCC) of testing cohort are 0.932, 0.923, 0.982, and 0.791, respectively. The nomogram has also been verified to have calibration curves with good overlap.ConclusionsRadiomics is beneficial in the malignancy prediction of BI-RADS 4 breast lesions. The radiomics predictive model built by the combination of DCE, DWI, and T2WI sequences has great application potential.

Comparison of artificial intelligence to the veterinary radiologist's diagnosis of canine cardiogenic pulmonary edema.

Application of artificial intelligence (AI) to improve clinical diagnosis is a burgeoning field in human and veterinary medicine. The objective of this prospective, diagnostic accuracy study was to determine the accuracy, sensitivity, and specificity of an AI-based software for diagnosing canine cardiogenic pulmonary edema from thoracic radiographs, using an American College of Veterinary Radiology-certified veterinary radiologist's interpretation as the reference standard. Five hundred consecutive canine thoracic radiographs made after-hours by a veterinary Emergency Department were retrieved. A total of 481 of 500 cases were technically analyzable. Based on the radiologist's assessment, 46 (10.4%) of these 481 dogs were diagnosed with cardiogenic pulmonary edema (CPE+). Of these cases, the AI software designated 42 of 46 as CPE+ and four of 46 as cardiogenic pulmonary edema negative (CPE-). Accuracy, sensitivity, and specificity of the AI-based software compared to radiologist diagnosis were 92.3%, 91.3%, and 92.4%, respectively (positive predictive value, 56%; negative predictive value, 99%). Findings supported using AI software screening for thoracic radiographs of dogs with suspected cardiogenic pulmonary edema to assist with short-term decision-making when a radiologist is unavailable.

Artificial intelligence for stepwise diagnosis and monitoring of COVID-19.

BackgroundMain challenges for COVID-19 include the lack of a rapid diagnostic test, a suitable tool to monitor and predict a patient’s clinical course and an efficient way for data sharing among multicenters. We thus developed a novel artificial intelligence system based on deep learning (DL) and federated learning (FL) for the diagnosis, monitoring, and prediction of a patient’s clinical course.MethodsCT imaging derived from 6 different multicenter cohorts were used for stepwise diagnostic algorithm to diagnose COVID-19, with or without clinical data. Patients with more than 3 consecutive CT images were trained for the monitoring algorithm. FL has been applied for decentralized refinement of independently built DL models.ResultsA total of 1,552,988 CT slices from 4804 patients were used. The model can diagnose COVID-19 based on CT alone with the AUC being 0.98 (95% CI 0.97–0.99), and outperforms the radiologist’s assessment. We have also successfully tested the incorporation of the DL diagnostic model with the FL framework. Its auto-segmentation analyses co-related well with those by radiologists and achieved a high Dice’s coefficient of 0.77. It can produce a predictive curve of a patient’s clinical course if serial CT assessments are available.InterpretationThe system has high consistency in diagnosing COVID-19 based on CT, with or without clinical data. Alternatively, it can be implemented on a FL platform, which would potentially encourage the data sharing in the future. It also can produce an objective predictive curve of a patient’s clinical course for visualization.Key Points• CoviDet could diagnose COVID-19 based on chest CT with high consistency; this outperformed the radiologist’s assessment. Its auto-segmentation analyses co-related well with those by radiologists and could potentially monitor and predict a patient’s clinical course if serial CT assessments are available. It can be integrated into the federated learning framework.• CoviDet can be used as an adjunct to aid clinicians with the CT diagnosis of COVID-19 and can potentially be used for disease monitoring; federated learning can potentially open opportunities for global collaboration.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00330-021-08334-6.

Is Perfusion Dose-Response Different Than Ventilation Dose-Response for Lung Cancer Patients Treated With Radiotherapy?

Ventilation-perfusion dose-response studies have focused on using functional imaging to characterize a dose-response and to assess whether the dose-response can predict subsequent pulmonary toxicity. Current dose-response studies typically focus on a single modality (ventilation or perfusion) and perform assessments on a population-based basis. There is no work evaluating whether ventilation dose-response differs from perfusion dose-response for individual lung cancer patients. The purpose of this study was to use single photon emission computed tomography (SPECT)-CT ventilation-perfusion imaging to quantitatively and clinically evaluate whether lung cancer patients treated with radiotherapy have differences between their ventilation and perfusion dose-responses.Twenty patients with biopsy proven lung cancer enrolled on a prospective functional avoidance protocol underwent SPECT-CT perfusion-ventilation scans pre- and 3-month post-radiotherapy. The SPECT-CT, planning CT, and dose distribution were rigidly registered. We calculated the relative changes in pre- to post-treatment ventilation-perfusion in lung regions receiving ≥ 20 Gy (referred to as ΔPV20 for perfusion and ΔVV20 for ventilation). Differences between perfusion-based ΔPV20 and ventilation-based ΔVV20 were assessed and referred to as ΔfV20. A radiologist read each pre- and post-treatment ventilation-perfusion scans using a semi-quantitative scale that assigned a score ranging from 1 to 8 that reflected functional defect and heterogeneity in each lung. The difference between radiologist scores for the pre- and post-treatment scans was calculated for both ventilation and perfusion. Differences between ventilation-based and perfusion-based radiologist scores were assessed.Twenty-five percent of patients had a difference of ΔfV20 > 35% between ventilation-perfusion pre- to post-treatment changes (ΔVV20 and ΔPV20) and 20% of patients had opposite directions for ventilation-perfusion pre- to post-treatment changes. Radiologist assessment resulted in 20% of patients having different ventilation-perfusion semi-quantitative change scores for ipsilateral lungs.The current work employed quantitative metrics and clinical assessments to evaluate whether ventilation dose-response differs from perfusion dose-response for individual lung cancer patients. Our data showed that ventilation dose-response can differ from perfusion dose-response for 20% to 30% of patients. Our study demonstrates that when evaluating functional dose-response for lung cancer patients, it is insufficient to look at ventilation or perfusion alone for a subset of patients; but rather both modes of functional imaging may be needed when predicting for clinical outcomes.

Chest imaging findings in COVID-19-positive patients in an Australian tertiary hospital.

IntroductionCoronavirus disease 2019 (COVID‐19) has infected over 215 million individuals worldwide. Chest radiographs (CXR) and computed tomography (CT) have assisted with diagnosis and assessment of COVID‐19. Previous reports have described peripheral and lower zone predominant opacities on chest radiographs. Whilst the most common patterns on CT are bilateral, peripheral basal predominant ground glass opacities (Wong et al., Radiology, 296, 2020, E72; Karimian and Azami, Pol J Radiol, 86, 2021, e31). This study describes the imaging findings in an Australian tertiary hospital population.MethodsCOVID‐PCR‐positive patients who had chest imaging (CXR, CT and ventilation perfusion (V/Q) scans) from January 2020 to August 2020 were included. Distribution, location and pattern of involvement was recorded. Evaluation of the assessors was performed using Fleiss Kappa calculations for review of radiographic findings and qualitative analysis of CT findings.ResultsA total of 681 studies (616 CXRs, 59 CTs, 6 V/Q) from 181 patients were reviewed. The most common chest radiograph finding was bilateral lower lobe predominant diffuse opacification and most common CT pattern being ground glass opacities. Of the CT imaging, 33 were CT Pulmonary Angiograms of which five demonstrated acute pulmonary emboli. There was good inter‐rater agreement between radiologists in assessment of imaging appearances on CXR (kappa 0.29–0.73) and CT studies.ConclusionA review of imaging in an Australian tertiary hospital demonstrates similar patterns of COVID‐19 infection on chest X‐ray and CT imaging when compared to the international population.

Diagnostic performance of qualitative and radiomics approach to parotid gland tumors: which is the added benefit of texture analysis?

Objective:To investigate whether MRI-based texture analysis improves diagnostic performance for the diagnosis of parotid gland tumors compared to conventional radiological approach.Methods:Patients with parotid gland tumors who underwent salivary glands MRI between 2008 and 2019 were retrospectively selected. MRI analysis included a qualitative assessment by two radiologists (one of which subspecialized on head and neck imaging), and texture analysis on various sequences. Diagnostic performances including sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) of qualitative features, radiologists’ diagnosis, and radiomic models were evaluated.Results:Final study cohort included 57 patients with 74 tumors (27 pleomorphic adenomas, 40 Warthin tumors, 8 malignant tumors). Sensitivity, specificity, and AUROC for the diagnosis of malignancy were 75%, 97% and 0.860 for non-subspecialized radiologist, 100%, 94% and 0.970 for subspecialized radiologist and 57.2%, 93.4%, and 0.927 using a MRI radiomics model obtained combining texture analysis on various MRI sequences. Sensitivity, specificity, and AUROC for the differential diagnosis between pleomorphic adenoma and Warthin tumors were 81.5%, 70%, and 0.757 for non-subspecialized radiologist, 81.5%, 95% and 0.882 for subspecialized radiologist and 70.8%, 82.5%, and 0.808 using a MRI radiomics model based on texture analysis of T2 weighted sequence. A combined radiomics model obtained with all MRI sequences yielded a sensitivity of 91.5% for the diagnosis of pleomorphic adenoma.Conclusion:MRI qualitative radiologist assessment outperforms radiomic analysis for the diagnosis of malignancy. MRI predictive radiomics models improves the diagnostic performance of non-subspecialized radiologist for the differential diagnosis between pleomorphic adenoma and Warthin tumor, achieving similar performance to the subspecialized radiologist.Advances in knowledge:Radiologists outperform radiomic analysis for the diagnosis of malignant parotid gland tumors, with some MRI qualitative features such as ill-defined margins, perineural spread, invasion of adjacent structures and enlarged lymph nodes being highly specific for malignancy. A radiomic model based on texture analysis of T2 weighted images yields higher specificity for the diagnosis of pleomorphic adenoma compared to a radiologist non-subspecialized in head and neck radiology, thus minimizing false-positive pleomorphic adenoma diagnosis rate and reducing unnecessary surgical complications.

CT Radiomic Features of Superior Mesenteric Artery Involvement in Pancreatic Ductal Adenocarcinoma: A Pilot Study.

Background Current imaging methods for prediction of complete margin resection (R0) in patients with pancreatic ductal adenocarcinoma (PDAC) are not reliable. Purpose To investigate whether tumor-related and perivascular CT radiomic features improve preoperative assessment of arterial involvement in patients with surgically proven PDAC. Materials and Methods This retrospective study included consecutive patients with PDAC who underwent surgery after preoperative CT between 2012 and 2019. A three-dimensional segmentation of PDAC and perivascular tissue surrounding the superior mesenteric artery (SMA) was performed on preoperative CT images with radiomic features extracted to characterize morphology, intensity, texture, and task-based spatial information. The reference standard was the pathologic SMA margin status of the surgical sample: SMA involved (tumor cells ≤1 mm from margin) versus SMA not involved (tumor cells >1 mm from margin). The preoperative assessment of SMA involvement by a fellowship-trained radiologist in multidisciplinary consensus was the comparison. High reproducibility (intraclass correlation coefficient, 0.7) and the Kolmogorov-Smirnov test were used to select features included in the logistic regression model. Results A total of 194 patients (median age, 66 years; interquartile range, 60-71 years; age range, 36-85 years; 99 men) were evaluated. Aside from surgery, 148 patients underwent neoadjuvant therapy. A total of 141 patients' samples did not involve SMA, whereas 53 involved SMA. A total of 1695 CT radiomic features were extracted. The model with five features (maximum hugging angle, maximum diameter, logarithm robust mean absolute deviation, minimum distance, square gray level co-occurrence matrix correlation) showed a better performance compared with the radiologist assessment (model vs radiologist area under the curve, 0.71 [95% CI: 0.62, 0.79] vs 0.54 [95% CI: 0.50, 0.59]; P < .001). The model showed a sensitivity of 62% (33 of 53 patients) (95% CI: 51, 77) and a specificity of 77% (108 of 141 patients) (95% CI: 60, 84). Conclusion A model based on tumor-related and perivascular CT radiomic features improved the detection of superior mesenteric artery involvement in patients with pancreatic ductal adenocarcinoma. © RSNA, 2021 Online supplemental material is available for this article. See also the editorial by Do and Kambadakone in this issue.

Evaluation of image quality and radiation dose with prospective ECG-gated 80 multi-slice CT scan in 182 consecutive children examinations with congenital heart disease

Computed tomography angiography (CTA) is a reliable imaging tool to evaluate children congenital heart disease (CHD) but requires radiation exposure. This study aims to investigate image quality and radiation dose with a prospective electrocardiogram (ECG)-gated 80 multi-slice CTA, in 182 consecutive children examinations with CHD. In total, 182 consecutive examinations performed with a prospective ECG-gated 80 multi-slice CTA from March 2016 to December 2017 in Nantes university hospital were retrospectively analyzed. Radiation dose was assessed by dose length product (DLP) and effective dose (E). Objective quality image was assessed by contrast to noise ratio (CNR). Subjective quality image was assessed by two radiologists with a 10-point scale evaluating visualization of coronary arteries. Correlation between quality image, radiation dose, age, weight, heart rate was analyzed. DLP and E were 29.3 ± 15.5 mGy.cm and 0.6 ± 0.2 mSv, respectively. Agreement between radiologists in assessment of subjective image quality was 0.79. Subjective image quality score was significantly better ( P < 0.01) with higher age, weight, DLP and lower heart rate in univariate analysis. CNR was significantly lower with increasing weight and DLP ( P < 0.05) but there was no correlation between CNR and age and heart rate. Prospective ECG-gated 80 multi-slice CTA is performing for the evaluation of children CHD with good image quality and low radiation dose. Visualization of small structures is better when age and weight increase, heart rate decreases and when radiation dose is more important. CNR evolves in the opposite side to subjective image quality score. It seems less appropriate to assess image quality from clinical viewpoint of the radiologist.