Inexperienced Readers Research Articles

Development and evaluation of machine-learning methods in whole-body magnetic resonance imaging with diffusion weighted imaging for staging of patients with cancer: the MALIBO diagnostic test accuracy study

Background Whole-body magnetic resonance imaging is accurate, efficient and cost-effective for cancer staging. Machine learning may support radiologists reading whole-body magnetic resonance imaging. Objectives To develop a machine-learning algorithm to detect normal organs and cancer lesions. To compare diagnostic accuracy, time and agreement of radiology reads to detect metastases using whole-body magnetic resonance imaging with concurrent machine learning (whole-body magnetic resonance imaging + machine learning) against standard whole-body magnetic resonance imaging (whole-body magnetic resonance imaging + standard deviation). Design and participants Retrospective analysis of (1) prospective single-centre study in healthy volunteers > 18 years (n = 51) and (2) prospective multicentre STREAMLINE study patient data (n = 438). Tests Index: whole-body magnetic resonance imaging + machine learning. Comparator: whole-body magnetic resonance imaging + standard deviation. Reference standard Previously established expert panel consensus reference at 12 months from diagnosis. Outcome measures Primary: difference in per-patient specificity between whole-body magnetic resonance imaging + machine learning and whole-body magnetic resonance imaging + standard deviation. Secondary: per-patient sensitivity, per-lesion sensitivity and specificity, read time and agreement. Methods Phase 1: classification forests, convolutional neural networks, and a multi-atlas approaches for organ segmentation. Phase 2/3: whole-body magnetic resonance imaging scans were allocated to Phase 2 (training = 226, validation = 45) and Phase 3 (testing = 193). Disease sites were manually labelled. The final algorithm was applied to 193 Phase 3 cases, generating probability heatmaps. Twenty-five radiologists (18 experienced, 7 inexperienced in whole-body magnetic resonance imaging) were randomly allocated whole-body magnetic resonance imaging + machine learning or whole-body magnetic resonance imaging + standard deviation over two or three rounds in a National Health Service setting. Read time was independently recorded. Results Phases 1 and 2: convolutional neural network had best Dice similarity coefficient, recall and precision measurements for healthy organ segmentation. Final algorithm used a ‘two-stage’ initial organ identification followed by lesion detection. Phase 3: evaluable scans (188/193, of which 50 had metastases from 117 colon, 71 lung cancer cases) were read between November 2019 and March 2020. For experienced readers, per-patient specificity for detection of metastases was 86.2% (whole-body magnetic resonance imaging + machine learning) and 87.7% (whole-body magnetic resonance imaging + standard deviation), (difference −1.5%, 95% confidence interval −6.4% to 3.5%; p = 0.387); per-patient sensitivity was 66.0% (whole-body magnetic resonance imaging + machine learning) and 70.0% (whole-body magnetic resonance imaging + standard deviation) (difference −4.0%, 95% confidence interval −13.5% to 5.5%; p = 0.344). For inexperienced readers (53 reads, 15 with metastases), per-patient specificity was 76.3% in both groups with sensitivities of 73.3% (whole-body magnetic resonance imaging + machine learning) and 60.0% (whole-body magnetic resonance imaging + standard deviation). Per-site specificity remained high within all sites; above 95% (experienced) or 90% (inexperienced). Per-site sensitivity was highly variable due to low number of lesions in each site. Reading time lowered under machine learning by 6.2% (95% confidence interval −22.8% to 10.0%). Read time was primarily influenced by read round with round 2 read times reduced by 32% (95% confidence interval 20.8% to 42.8%) overall with subsequent regression analysis showing a significant effect (p = 0.0281) by using machine learning in round 2 estimated as 286 seconds (or 11%) quicker. Interobserver variance for experienced readers suggests moderate agreement, Cohen’s κ = 0.64, 95% confidence interval 0.47 to 0.81 (whole-body magnetic resonance imaging + machine learning) and Cohen’s κ = 0.66, 95% confidence interval 0.47 to 0.81 (whole-body magnetic resonance imaging + standard deviation). Limitations Patient whole-body magnetic resonance imaging data were heterogeneous with relatively few metastatic lesions in a wide variety of locations, making training and testing difficult and hampering evaluation of sensitivity. Conclusions There was no difference in diagnostic accuracy for whole-body magnetic resonance imaging radiology reads with or without machine-learning support, although radiology read time may be slightly shortened using whole-body magnetic resonance imaging + machine learning. Future work Failure-case analysis to improve model training, automate lesion segmentation and transfer of machine-learning techniques to other tumour types and imaging modalities. Study registration This study is registered as ISRCTN23068310. Funding This award was funded by the National Institute for Health and Care Research (NIHR) Efficacy and Mechanism Evaluation (EME) programme (NIHR award ref: 13/122/01) and is published in full in Efficacy and Mechanism Evaluation; Vol. 11, No. 15. See the NIHR Funding and Awards website for further award information.

Effect of online training on the reliability of assessing sacroiliac joint radiographs in axial spondyloarthritis - a randomized, controlled study.

Radiographic assessment of sacroiliac joints (SIJs) according to the modified New York (mNY) criteria is key in classification of axSpA but has moderate inter-reader agreement. We aimed to investigate the reliability improvements scoring SIJ radiographs after applying an online real-time iterative calibration (RETIC) module, in addition to a slideshow and video alone. Nineteen readers, randomized to 2 groups (A/B), completed 3 calibration steps: I) review of manuscripts, II) review of slideshow and video and group A completed RETIC, III) re-review of slideshow+video and group B completed RETIC . The RETIC module gave instant feedback on reader's gradings and continued until predefined reliability (kappa) targets for mNY positivity/negativity were met. Each step was followed by scoring different batches of 25 radiographs (Exercises I-III). Agreement (kappa) with an expert radiologist was assessed for mNY+/mNY- and individual lesions. Improvements by training strategies were tested by linear mixed models. In exercises I/II/III, mNY kappas were 0.61/0.76/0.84 in group A, and 0.70/0.68/0.86 in group B, respectively, i.e. increasing, mainly after RETIC completion. Improvements were observed for both grading mNY+/mNY- and for individual pathologies, both in experienced and, particularly, inexperienced readers. Completion of the RETIC module in addition to slideshow and video caused a significant kappa increase of 0.17 (CI: 0.07-0.27, P=0.002) for mNY+/mNY- grading, while completion of slideshow and video alone did not (0.0; CI -0.10-+0.10, P=0.99). Agreement on scoring radiographs according to the mNY criteria significantly improved when adding an online RETIC module, but not by slideshow and video alone.

Impact of an expert-derived, quick hands-on tool on classifying pulmonary hypertension in chest computed tomography: a study on inexperienced readers using RAPID-CT-PH

PurposeTo test the inter-reader agreement in classifying pulmonary hypertension (PH) on chest contrast-enhanced computed tomography (CECT) between a consensus of two cardio-pulmonary-devoted radiologists (CRc) and inexperienced readers (radiology residents, RRs) when using a CECT-based quick hands-on tool built upon PH imaging literature, i.e., the “Rapid Access and Practical Information Digest on Computed Tomography for PH-RAPID-CT-PH”.Material and methodsThe observational study retrospectively included 60 PH patients who underwent CECT between 2015 and 2022. Four RRs independently reviewed all CECTs and classified each case into one of the five PH groups per the 2022 ESC/ERS guidelines. While RR3 and RR4 (RAPID-CT-PH group) used RAPID-CT-PH, RR1 and RR2 (control group) did not. RAPID-CT-PH and control groups’ reports were compared with CRc using unweighted Cohen’s Kappa (k) statistics. RRs’ report completeness and reporting time were also compared using the Wilcoxon–Mann–Whitney test.ResultsThe inter-reader agreement in classifying PH between the RAPID-CT-PH group and CRc was substantial (k = 0.75 for RR3 and k = 0.65 for RR4); while, it was only moderate for the control group (k = 0.57 for RR1 and k = 0.49 for RR2). Using RAPID-CT-PH resulted in significantly higher report completeness (all p < 0.0001) and significantly lower reporting time (p < 0.0001) compared to the control group.ConclusionRRs using RAPID-CT-PH showed a substantial agreement with CRc on CECT-based PH classification. RAPID-CT-PH improved report completeness and reduced reporting time. A quick hands-on tool for classifying PH on chest CECT may help inexperienced radiologists effectively contribute to the PH multidisciplinary team.

Improving the diagnostic performance of inexperienced readers for thyroid nodules through digital self-learning and artificial intelligence assistance.

Data-driven digital learning could improve the diagnostic performance of novice students for thyroid nodules. To evaluate the efficacy of digital self-learning and artificial intelligence-based computer-assisted diagnosis (AI-CAD) for inexperienced readers to diagnose thyroid nodules. Between February and August 2023, a total of 26 readers (less than 1 year of experience in thyroid US from various departments) from 6 hospitals participated in this study. Readers completed an online learning session comprising 3,000 thyroid nodules annotated as benign or malignant independently. They were asked to assess a test set consisting of 120 thyroid nodules with known surgical pathology before and after a learning session. Then, they referred to AI-CAD and made their final decisions on the thyroid nodules. Diagnostic performances before and after self-training and with AI-CAD assistance were evaluated and compared between radiology residents and readers from different specialties. AUC (area under the receiver operating characteristic curve) improved after the self-learning session, and it improved further after radiologists referred to AI-CAD (0.679 vs 0.713 vs 0.758, p<0.05). Although the 18 radiology residents showed improved AUC (0.7 to 0.743, p=0.016) and accuracy (69.9% to 74.2%, p=0.013) after self-learning, the readers from other departments did not. With AI-CAD assistance, sensitivity (radiology 70.3% to 74.9%, others 67.9% to 82.3%, all p<0.05) and accuracy (radiology 74.2% to 77.1%, others 64.4% to 72.8%, all p <0.05) improved in all readers. While AI-CAD assistance helps improve the diagnostic performance of all inexperienced readers for thyroid nodules, self-learning was only effective for radiology residents with more background knowledge of ultrasonography. Online self-learning, along with AI-CAD assistance, can effectively enhance the diagnostic performance of radiology residents in thyroid cancer.

Validation of the standardization framework SSTR-RADS 1.0 for neuroendocrine tumors using the novel SSTR‑targeting peptide [18F]SiTATE

ObjectivesSomatostatin receptor positron emission tomography/computed tomography (SSTR-PET/CT) using [68Ga]-labeled tracers is a widely used imaging modality for neuroendocrine tumors (NET). Recently, [18F]SiTATE, a SiFAlin tagged [Tyr3]-octreotate (TATE) PET tracer, has shown great potential due to favorable clinical characteristics. We aimed to evaluate the reproducibility of Somatostatin Receptor-Reporting and Data System 1.0 (SSTR-RADS 1.0) for structured interpretation and treatment planning of NET using [18F]SiTATE.MethodsFour readers assessed [18F]SiTATE-PET/CT of 95 patients according to the SSTR-RADS 1.0 criteria at two different time points. Each reader evaluated up to five target lesions per scan. The overall scan score and the decision on peptide receptor radionuclide therapy (PRRT) were considered. Inter- and intra-reader agreement was determined using the intraclass correlation coefficient (ICC).ResultsThe ICC analysis on the inter-reader agreement using SSTR-RADS 1.0 for identical target lesions (ICC ≥ 85%), overall scan score (ICC ≥ 90%), and the decision to recommend PRRT (ICC ≥ 85%) showed excellent agreement. However, significant differences were observed in recommending PRRT among experienced readers (ER) (p = 0.020) and inexperienced readers (IR) (p = 0.004). Compartment-based analysis demonstrated good to excellent inter-reader agreement for most organs (ICC ≥ 74%), except for lymph nodes (ICC ≥ 53%).ConclusionSSTR-RADS 1.0 represents a highly reproducible and consistent framework system for stratifying SSTR-targeted PET/CT scans, even using the novel SSTR-ligand [18F]SiTATE. Some inter-reader variability was observed regarding the evaluation of uptake intensity prior to PRRT as well as compartment scoring of lymph nodes, indicating that those categories require special attention during further clinical validation and might be refined in a future SSTR-RADS version 1.1.Clinical relevance statementSSTR-RADS 1.0 is a consistent framework for categorizing somatostatin receptor-targeted PET/CT scans when using [18F]SiTATE. The framework serves as a valuable tool for facilitating and improving the management of patients with NET.Key PointsSSTR-RADS 1.0 is a valuable tool for managing patients with NET.SSTR-RADS 1.0 categorizes patients with showing strong agreement across diverse reader expertise.As an alternative to [68Ga]-labeled PET/CT in neuroendocrine tumor imaging, SSTR-RADS 1.0 reliably classifies [18F]SiTATE-PET/CT.

Human-Level Differentiation of Medulloblastoma from Pilocytic Astrocytoma: A Real-World Multicenter Pilot Study.

Medulloblastoma and pilocytic astrocytoma are the two most common pediatric brain tumors with overlapping imaging features. In this proof-of-concept study, we investigated using a deep learning classifier trained on a multicenter data set to differentiate these tumor types. We developed a patch-based 3D-DenseNet classifier, utilizing automated tumor segmentation. Given the heterogeneity of imaging data (and available sequences), we used all individually available preoperative imaging sequences to make the model robust to varying input. We compared the classifier to diagnostic assessments by five readers with varying experience in pediatric brain tumors. Overall, we included 195 preoperative MRIs from children with medulloblastoma (n = 69) or pilocytic astrocytoma (n = 126) across six university hospitals. In the 64-patient test set, the DenseNet classifier achieved a high AUC of 0.986, correctly predicting 62/64 (97%) diagnoses. It misclassified one case of each tumor type. Human reader accuracy ranged from 100% (expert neuroradiologist) to 80% (resident). The classifier performed significantly better than relatively inexperienced readers (p < 0.05) and was on par with pediatric neuro-oncology experts. Our proof-of-concept study demonstrates a deep learning model based on automated tumor segmentation that can reliably preoperatively differentiate between medulloblastoma and pilocytic astrocytoma, even in heterogeneous data.

Usar una herramienta comercial de inteligencia artificial no entrenada para COVID-19 mejora ligeramente la interpretación de las radiografías con neumonía COVID-19, especialmente entre lectores inexpertos

IntroductionOur objective is to evaluate how useful an artificial intelligence (AI) tool is to chest radiograph readers with various levels of expertise for the diagnosis of COVID-19 pneumonia when the tool has been trained on a non-COVID-19 pneumonia pathology. MethodsData was collected for patients who had previously undergone a chest radiograph and digital tomosynthesis due to suspected COVID-19 pneumonia. The gold standard consisted of the readings of two expert radiologists who assessed the presence and distribution of COVID-19 pneumonia on the images. Six medical students, two radiology trainees, and two other expert thoracic radiologists participated as additional readers. Two radiograph readings and a third supported by the AI Thoracic Care Suite tool were performed. COVID-19 pneumonia distribution and probability were assessed along with the contribution made by AI. Agreement and diagnostic performance were analysed. ResultsThe sample consisted of 113 cases, of which 56 displayed lung opacities, 52.2% were female, and the mean age was 50.70±14.9. Agreement with the gold standard differed between students, trainees, and radiologists. There was a non-significant improvement for four of the six students when AI was used. The use of AI by students did not improve the COVID-19 pneumonia diagnostic performance but it did reduce the difference in diagnostic performance with the more expert radiologists. Furthermore, it had more influence on the interpretation of mild pneumonia than severe pneumonia and normal radiograph findings. AI resolved more doubts than it generated, especially among students (31.30 vs 8.32%), followed by trainees (14.45 vs 5.7%) and radiologists (10.05% vs 6.15%). ConclusionFor expert and lesser experienced radiologists, this commercial AI tool has shown no impact on chest radiograph readings of patients with suspected COVID-19 pneumonia. However, it aided the assessment of inexperienced readers and in cases of mild pneumonia.

한국 프랑스어 전공 대학학습자의 프랑스어 인터넷 신문기사의 장르파악을 위한 읽기 활동 고찰

Objectives This study aimed to investigate the communicative competence and reading strategies employed by Korean university learners majoring in French in actual reading activities, with the overarching goal of enhancing communicative competence through reading education.&#x0D; Methods To achieve this, the study focused on reading activities related to genre identification in French online newspaper articles. Ten Korean university students majoring in French and six native speakers participated in the reading experiment. The study utilized an eye-tracker and specifically analyzed fixation times (gaze duration, re-reading, and total reading time) required when reading words of the section and words indicating publication dates, based on information obtained from interviews.&#x0D; Results Participants, driven by the reading goal of genre identification, selectively identified and integrated spatially dispersed words and phrases in the text, understanding the genre through their different interactions. Fixation times required to read words of the section and words indicating publication dates revealed differences between native speakers and Korean participants. In the case of native readers, who were considered experienced readers, fixation times were longer for words of the section, regardless of syllabic length, referential competence, or reading path. On the other hand, for Korean participants who were considered inexperienced readers in French reading, fixation times were influenced by syllabic length, comprehension, and reading path.&#x0D; Conclusions The study highlighted that participants utilized communicative competence and selection strategies in reading activities aimed at genre identification. However, differences in reading aspects between the two groups indicated a particular deficiency in discourse competence among Korean participants in terms of communicative competence. In response, a pedagogical strategy that combines the teaching method ‘global approach to written texts' and Findlay and Walker's theory, an eye movement control model, was proposed. The research confirmed that understanding reading activity aspects is a meaningful approach to examining the (foreign language) communicative competence used by learners. It also emphasized the need for active reading that goes beyond relying solely on text, considering the characteristics of (non)linguistic elements provided by the text.&#x0D;

Assessing the impact of structured reporting on learning how to report lung cancer staging CT: A triple cohort study on inexperienced readers

PurposeTo assess the clinical utility of chest computed tomography (CT) reports for non-small-cell lung cancer (NSCLC) staging generated by inexperienced readers using structured reporting (SR) templates from the Royal College of Radiologists (RCR-SR) and the Italian Society of Medical and Interventional Radiology (SIRM-SR), compared to traditional non-systematic reports (NSR). MethodsIn a cohort of 30 NSCLC patients, six third-year radiology residents reported CT examinations in two 2-month-apart separate sessions using NSR in the first and NSR, RCR-SR, or SIRM-SR in the second. Couples of expert radiologists and thoracic oncologists in consensus evaluated completeness, accuracy, and clarity. All the quality indicators were expressed on a 100-point scale. The Wilcoxon signed ranks, and Wilcoxon-Mann Whitney tests were used for statistical analyses. ResultsResults showed significantly higher completeness for RCR-SR (90 %) and SIRM-SR (100 %) compared to NSR (70 %) in the second session (all p < 0.001). SIRM-SR demonstrated superior accuracy (70 % vs. 55 %, p < 0.001) over NSR, while RCR-SR and NSR accuracy did not significantly differ (60 % vs. 62.5 %, p = 0.06). In the second session, RCR-SR and SIRM-SR surpassed NSR in completeness, accuracy, and clarity (all p < 0.001, except p = 0.04 for accuracy between RCR-SR and NSR). SIRM-SR outperformed RCR-SR in completeness (100 % vs. 90 %, p < 0.001) and accuracy (70 % vs. 62.5 %, p = 0.002), with equivalent clarity (90 % for both, p = 0.27). ConclusionsInexperienced readers using RCR-SR and SIRM-SR demonstrated high-quality reporting, indicating their potential in radiology residency programs to enhance reporting skills for NSCLC staging and effective interaction with all the physicians involved in managing NSCLC patients.

Integrating radiomics with the vesical imaging-reporting and data system to predict muscle invasion of bladder cancer.

To develop predictive models based on the integration of radiomics with the Vesical Imaging-Reporting and Data System (VI-RADS) for determining muscle invasion of bladder cancer. One hundred ninety-one patients were retrospectively included in this study from January 2015 to March 2022. Of these, 121 data were randomly divided into training and validation sets at a ratio of 7:3. The remaining data (n = 70) served as the independent testing set. The radiomics features were extracted from bladder cancer on high-b-value DWI images. The pipelines of radiomics models were trained in the training set. One optimal model was selected based on the performance in the validation set. Then, the selected model was tested in the independent testing set. Two radiologists evaluated the VI-RADS based on T2WI and DWI. Reader 1 was an experienced reader, and Reader 2 was an inexperienced reader. A clinical-radiomics model was built by integrating the radiomics signature and VI-RADS. The performance was assessed using receiver operating characteristic curve analysis. The histopathological results were used as the standard reference to assess the diagnostic accuracy of muscle invasion. The radiomics model had area under the curve (AUC) values of 0.801, 0.867, and 0.806 in the training, validation, and testing sets, respectively. The VI-RADS scores of Readers 1/2 yielded AUC values of 0.831/0.781, 0.909/0.815, and 0.871/0.776 in the training, validation, and testing sets, respectively. The clinical-radiomics model for Readers 1/2 revealed AUC values of 0.889/0.854, 0.961/0.919, and 0.881/0.844 in the training, validation, and testing sets, respectively. The performance of the clinical-radiomics model was improved compared to the VI-RADS score for inexperienced Reader 2 (P < 0.05). The radiomics model was useful in the diagnosis of muscle invasion of bladder cancer. The clinical-radiomics model integrating radiomics and VI-RADS further improved the performance compared to VI-RADS alone, which was helpful for readers with less diagnostic experience.

Genetic manipulation and targeted protein degradation in mammalian systems: practical considerations, tips and tricks for discovery research.

Gaining a mechanistic understanding of the molecular pathways underpinning cellular and organismal physiology invariably relies on the perturbation of an experimental system to infer causality. This can be achieved either by genetic manipulation or by pharmacological treatment. Generally, the former approach is applicable to a wider range of targets, is more precise, and can address more nuanced functional aspects. Despite such apparent advantages, genetic manipulation (i.e., knock-down, knock-out, mutation, and tagging) in mammalian systems can be challenging due to problems with delivery, low rates of homologous recombination, and epigenetic silencing. The advent of CRISPR-Cas9 in combination with the development of robust differentiation protocols that can efficiently generate a variety of different cell types in vitro has accelerated our ability to probe gene function in a more physiological setting. Often, the main obstacle in this path of enquiry is to achieve the desired genetic modification. In this short review, we will focus on gene perturbation in mammalian cells and how editing and differentiation of pluripotent stem cells can complement more traditional approaches. Additionally, we introduce novel targeted protein degradation approaches as an alternative to DNA/RNA-based manipulation. Our aim is to present a broad overview of recent approaches and in vitro systems to study mammalian cell biology. Due to space limitations, we limit ourselves to providing the inexperienced reader with a conceptual framework on how to use these tools, and for more in-depth information, we will provide specific references throughout.

Orthographic and phonological processing effects on the reading abilities of young children learning to read Malayalam alphasyllabary

Background Learning to read is a complex process that involves phonological and orthographic processing abilities, broader language skills and cognitive processes across all writing systems. Although these components remain common, the pace of acquisition of phonological and orthographic processing and reading abilities differ across writing systems. Methods This study aimed to understand the developmental pattern of reading and phonological processing, assess their interrelationship and study the predictors of reading ability among 175 children, learning to read Malayalam and attending Grades 1–3. They completed various reading (akshara reading, word and nonword reading) and phonological processing (rhyme recognition, syllable deletion, syllable segmentation, phoneme deletion, phoneme substitution, rapid naming and nonword repetition) tasks. Results Grade and reading level-based analysis was performed to collate the results. Results revealed a significant overlap of phonological processing and reading abilities among children in Grades 2 and 3. Reading scores demonstrated a weak-to-moderate significant correlation with the phonological processing tasks. Although skilled readers performed significantly better on all phonological processing tasks, their scores did not reach ceiling. Syllable-level tasks were easier compared to phoneme-level tasks, and they significantly predicted the reading abilities of children. Conclusions The findings contribute to the existing body of evidence for akshara writing systems, specifically in Malayalam alpha-syllabary. The development of phonological processing and reading in Malayalam extends beyond the third grade, and syllable processing contributed to increased variance in reading accuracy among inexperienced readers.

Diagnostic performance of deep learning-based vessel extraction and stenosis detection on coronary computed tomography angiography for coronary arterydisease: a multi-reader multi-case study.

Post-processing and interpretation of coronary CT angiography (CCTA) imaging are time-consuming and dependent on the reader's experience. An automated deep learning (DL)-based imaging reconstruction and diagnosis system was developed to improve diagnostic accuracy and efficiency. Our study including 374 cases from five sites, inviting 12 radiologists, assessed the DL-based system in diagnosing obstructive coronary disease with regard to diagnostic performance, imaging post-processing and reporting time of radiologists, with invasive coronary angiography as a standard reference. The diagnostic performance of DL system and DL-assisted human readers was compared with the traditional method of human readers without DL system. Comparing the diagnostic performance of human readers without DL system versus with DL system, the AUC was improved from 0.81 to 0.82 (p < 0.05) at patient level and from 0.79 to 0.81 (p < 0.05) at vessel level. An increase in AUC was observed in inexperienced radiologists (p < 0.05), but was absent in experienced radiologists. Regarding diagnostic efficiency, comparing the DL system versus human reader, the average post-processing and reporting time was decreased from 798.60s to 189.12s (p < 0.05). The sensitivity and specificity of using DL system alone were 93.55% and 59.57% at patient level and 83.23% and 79.97% at vessel level, respectively. With the DL system serving as a concurrent reader, the overall post-processing and reading time was substantially reduced. The diagnostic accuracy of human readers, especially for inexperienced readers, was improved. DL-assisted human reader had the potential of being the reading mode of choice in clinical routine.

Observer Sensitivity for Detection of Pulmonary Nodules in Ultra-Low Dose Computed Tomography Protocols Using a Third-Generation Dual-Source CT with Ultra-High Pitch—A Phantom Study

This study evaluates ultra-low-dose computed tomography (ULDCT) protocols concerning the detectability of pulmonary nodules. The influence of tube current settings, kernels, strength levels of third-generation iterative reconstruction algorithms, and pitch was investigated. A chest phantom with artificial spherical nodules of different densities and diameters was examined with a third-generation dual-source CT. Scanning and post-processing protocols, tube current levels, and ultra-high and non-high pitch modes were applied. Images were reconstructed with filtered back-projection (FBP) or advanced model-based iterative reconstruction (ADMIRE) algorithms. Sharp (Bl57) or medium-soft (Br36) convolution kernels were applied. The reading was performed by an experienced and an inexperienced reader. The highest observer sensitivity was found using a non-high pitch protocol at tube currents of 120 mAs and 90 mAs with the sharp kernel and iterative reconstruction level of 5. Non-high pitch protocols showed better detectability of solid nodules. Combinations with the medium-soft kernel achieved slightly higher observer sensitivity than with the sharp kernel. False positives (FP) occurred more often for subsolid nodules, at a tube current level of 120 mAs, and with the sharp kernel. A tube current level of 90 mAs combined with the highest iterative reconstruction level achieved the highest accuracy in lung nodule detection regardless of size, density, and reader experience.

Development and performance evaluation of a deep learning lung nodule detection system

BackgroundLung cancer is the leading cause of cancer-related deaths throughout the world. Chest computed tomography (CT) is now widely used in the screening and diagnosis of lung cancer due to its effectiveness. Radiologists must identify each small nodule shadow from 3D volume images, which is very burdensome and often results in missed nodules. To address these challenges, we developed a computer-aided detection (CAD) system that automatically detects lung nodules in CT images.MethodsA total of 1997 chest CT scans were collected for algorithm development. The algorithm was designed using deep learning technology. In addition to evaluating detection performance on various public datasets, its robustness to changes in radiation dose was assessed by a phantom study. To investigate the clinical usefulness of the CAD system, a reader study was conducted with 10 doctors, including inexperienced and expert readers. This study investigated whether the use of the CAD as a second reader could prevent nodular lesions in lungs that require follow-up examinations from being overlooked. Analysis was performed using the Jackknife Free-Response Receiver-Operating Characteristic (JAFROC).ResultsThe CAD system achieved sensitivity of 0.98/0.96 at 3.1/7.25 false positives per case on two public datasets. Sensitivity did not change within the range of practical doses for a study using a phantom. A second reader study showed that the use of this system significantly improved the detection ability of nodules that could be picked up clinically (p = 0.026).ConclusionsWe developed a deep learning-based CAD system that is robust to imaging conditions. Using this system as a second reader increased detection performance.

Interobserver Variability Between Expert, Experienced, and Novice Operator Affects Interpretation of Optical Coherence Tomography and 20 MHz Intravascular Ultrasound Imaging

BackgroundVessel sizing and calcium detection by intracoronary imaging informs optimal strategy during coronary interventions, but image interpretation and analysis software vary considerably between platforms. We compared the interobserver variability of clinicians with a range of experience in assessing co-registered optical coherence tomography (OCT) and 20 MHz solid state intravascular ultrasound (IVUS) vessel/ lumen geometry and quantitative plaque data. MethodsCo-registered OCT and IVUS frames at the minimum lumen area (MLA) and 5 frames at 2 mm intervals upstream and downstream were read blinded by an expert, consultant, interventional fellow and registrar to define vessel and lumen sizes, plaque characteristics (arc of calcium and lipid) and presence of OCT-defined thin-capped fibroatheroma (TCFA). ResultsOverall, 143 paired frames of OCT and IVUS were analysed. Excellent consistency was seen for all OCT measures of vessel/luminal geometry irrespective of experience (all intraclass correlation coefficients (ICC) >0.89). Inexperience compromised ICC for IVUS (lumen area ICC 0.56; vessel size ICC 0.65) and overestimated lumen size compared to expert (p < 0.001). Calcium arc agreement for OCT and IVUS was equally strong for all but the most inexperienced reader (ICC >0.84). OCT lipid arc agreement was moderate for experienced readers but poor for all grades of reader with IVUS (ICC 0.24–0.43). OCT-TCFA agreement was moderate between expert and consultant (κ = 0.55) and poor for less experienced readers. ConclusionOCT dimensions are minimally affected by observer experience and more consistent than IVUS. Inexperienced readers oversize with IVUS. Calcium arc is dependably assessed by IVUS and OCT by all but the most inexperienced reader.

Biparametric prostate MRI: impact of a deep learning-based software and of quantitative ADC values on the inter-reader agreement of experienced and inexperienced readers

ObjectiveTo investigate the impact of an artificial intelligence (AI) software and quantitative ADC (qADC) on the inter-reader agreement, diagnostic performance, and reporting times of prostate biparametric MRI (bpMRI) for experienced and inexperienced readers.Materials and methodsA total of 170 multiparametric MRI (mpMRI) of patients with suspicion of prostate cancer (PCa) were retrospectively reviewed by one experienced and one inexperienced reader three times, following a wash-out period. First, only the bpMRI sequences, including T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI) sequences, and apparent diffusion coefficient (ADC) maps, were used. Then, bpMRI and quantitative ADC values were used. Lastly, bpMRI and the AI software were used. Inter-reader agreement between the two readers and between each reader and the mpMRI original reports was calculated. Detection rates and reporting times were calculated for each group.ResultsInter-reader agreement with respect to mpMRI was moderate for bpMRI, Quantib, and qADC for both the inexperienced (weighted k of 0.42, 0.45, and 0.41, respectively) and the experienced radiologists (weighted k of 0.44, 0.46, and 0.42, respectively). Detection rate of PCa was similar between the inexperienced (0.24, 0.26, and 0.23) and the experienced reader (0.26, 0.27 and 0.27), for bpMRI, Quantib, and qADC, respectively. Reporting times were lower for Quantib (8.23, 7.11, and 9.87 min for the inexperienced reader and 5.62, 5.07, and 6.21 min for the experienced reader, for bpMRI, Quantib, and qADC, respectively).ConclusionsAI and qADC did not have a significant impact on the diagnostic performance of both readers. The use of Quantib was associated with lower reporting times.

Ultra-low dose CT for scaphoid fracture detection-a simulational approach to quantify the capability of radiation exposure reduction without diagnostic limitation.

BackgroundModern CT might deliver higher image quality than necessary for fracture imaging, which would mean non-essential effective radiation exposure for patients. We simulated ultra-low dose (ULD)-CT at different dose levels and analyzed their diagnostic performance for scaphoid fracture detection.Methods30 consecutive high quality CT with clinically suspected scaphoid fractures were assessed. ULD-simulations were made at 20%, 10% and 5% of original dose. Three readers at different levels of experience (expert, moderate, inexperienced) expressed their diagnostic confidence (DC; 5-point-Likert-scale) and analyzed the presence and classification of scaphoid fractures within Krimmer’s and Herbert’s classifications. Effective radiation exposure of the original data sets and ULD-CT were calculated.ResultsAt 20% and 10% dose the more experienced readers reached perfect sensitivity (100%) and specificity (100%), showing perfect agreement regarding fracture classification (1.00). Diagnostic performance decreased at 5% dose (92.86% sensitivity, 100% specificity; expert reader). The inexperienced reader showed reduced sensitivity and specificity at all dose levels. At 10% dose minimal DC of all readers was 3/5 and mean calculated effective radiation exposure was 1.11 [±0.36] µSv.ConclusionsThe results suggest that ULD-CT at 10% dose compared to high quality CT might offer sufficient image quality to precisely detect and classify scaphoid fractures, if moderate experience of the radiologist is granted.

Automated Identification of Coronary Arteries in Assisting Inexperienced Readers: Comparison between Two Commercial Vendors.

Background: to assess the performance and speed of two commercially available advanced cardiac software packages in the automated identification of coronary vessels as an aiding tool for inexperienced readers. Methods: Hundred and sixty patients undergoing coronary CT angiography (CCTA) were prospectively enrolled from February until September 2021 and randomized in two groups, each one composed by 80 patients. Patients in group 1 were scanned on Revolution EVO CT Scanner (GE Healthcare), while patients in group 2 had the CCTA performed on Brilliance iCT (Philips Healthcare); each examination was evaluated on the respective vendor proprietary advanced cardiac software (software 1 and 2, respectively). Two inexperienced readers in cardiac imaging verified the software performance in the automated identification of the three major coronary vessels: (RCA, LCx, and LAD) and in the number of identified coronary segments. Time of analysis was also recorded. Results: software 1 correctly and automatically nominated 202/240 (84.2%) of the three main coronary vessels, while software 2 correctly identified 191/240 (79.6%) (p = 0.191). Software 1 achieved greater performances in recognizing the LCx (81.2% versus 67.5%; p = 0.048), while no differences have been reported in detecting the RCA (p = 0.679), and the LAD (p = 0.618). On a per-segment analysis, software 1 outperformed software 2, automatically detecting 942/1062 (88.7%) coronary segments, while software 2 detected 797/1078 (73.9%) (p < 0.001). Average reconstruction and detection time was of 13.8 s for software 1 and 21.9 s for software 2 (p < 0.001). Conclusions: automated cardiac software packages are a reliable and time-saving tool for inexperienced reader. Software 1 outperforms software 2 and might therefore better assist inexperienced CCTA readers in automated identification of the three main vessels and coronaries segments, with a consistent time saving of the reading session.

The effect of CT texture-based analysis using machine learning approaches on radiologists' performance in differentiating focal-type autoimmune pancreatitis and pancreatic duct carcinoma

PurposeTo develop a support vector machine (SVM) classifier using CT texture-based analysis in differentiating focal-type autoimmune pancreatitis (AIP) and pancreatic duct carcinoma (PD), and to assess the radiologists’ diagnostic performance with or without SVM.Materials and methodsThis retrospective study included 50 patients (20 patients with focal-type AIP and 30 patients with PD) who underwent dynamic contrast-enhanced CT. Sixty-two CT texture-based features were extracted from 2D images of the arterial and portal phase CTs. We conducted data compression and feature selections using principal component analysis (PCA) and produced the SVM classifier. Four readers participated in this observer performance study and the statistical significance of differences with and without the SVM was assessed by receiver operating characteristic (ROC) analysis.ResultsThe SVM performance indicated a high performance in differentiating focal-type AIP and PD (AUC = 0.920). The AUC for all 4 readers increased significantly from 0.827 to 0.911 when using the SVM outputs (p = 0.010). The AUC for inexperienced readers increased significantly from 0.781 to 0.905 when using the SVM outputs (p = 0.310). The AUC for experienced readers increased from 0.875 to 0.912 when using the SVM outputs, however, there was no significant difference (p = 0.018).ConclusionThe use of SVM classifier using CT texture-based features improved the diagnostic performance for differentiating focal-type AIP and PD on CT.