Board-certified Radiologists Research Articles

Response accuracy of GPT-4 across languages: insights from an expert-level diagnostic radiology examination in Japan.

This study aims to investigate the effects of language selection and translation quality on Generative Pre-trained Transformer-4 (GPT-4)'s response accuracy to expert-level diagnostic radiology questions. We analyzed 146 diagnostic radiology questions from the Japan Radiology Board Examination (2020-2022), with consensus answers provided by two board-certified radiologists. The questions, originally in Japanese, were translated into English by GPT-4 and DeepL and into German and Chinese by GPT-4. Responses were generated by GPT-4 five times per question set per language. Response accuracy was compared between languages using one-way ANOVA with Bonferroni correction or the Mann-Whitney U test. Scores on selected English questions translated by a professional service and GPT-4 were also compared. The impact of translation quality on GPT-4's performance was assessed by linear regression analysis. The median scores (interquartile range) for the 146 questions were 70 (68-72) (Japanese), 89 (84.5-95.5) (GPT-4 English), 64 (55.5-67) (Chinese), and 56 (46.5-67.5) (German). Significant differences were found between Japanese and English (p = 0.002) and between Japanese and German (p = 0.022). The counts of correct responses across five attempts for each question were significantly associated with the quality of translation into English (GPT-4, DeepL) and German (GPT-4). In a subset of 31 questions where English translations yielded fewer correct responses than Japanese originals, professionally translated questions yielded better scores than those translated by GPT-4 (13 versus 8 points, p = 0.0079). GPT-4 exhibits higher accuracy when responding to English-translated questions compared to original Japanese questions, a trend not observed with German or Chinese translations. Accuracy improves with higher-quality English translations, underscoring the importance of high-quality translations in improving GPT-4's response accuracy to diagnostic radiology questions in non-English languages and aiding non-native English speakers in obtaining accurate answers from large language models.

TECRR: a benchmark dataset of radiological reports for BI-RADS classification with machine learning, deep learning, and large language model baselines

BackgroundRecently, machine learning (ML), deep learning (DL), and natural language processing (NLP) have provided promising results in the free-form radiological reports’ classification in the respective medical domain. In order to classify radiological reports properly, a high-quality annotated and curated dataset is required. Currently, no publicly available breast imaging-based radiological dataset exists for the classification of Breast Imaging Reporting and Data System (BI-RADS) categories and breast density scores, as characterized by the American College of Radiology (ACR). To tackle this problem, we construct and annotate a breast imaging-based radiological reports dataset and its benchmark results. The dataset was originally in Spanish. Board-certified radiologists collected and annotated it according to the BI-RADS lexicon and categories at the Breast Radiology department, TecSalud Hospitals Monterrey, Mexico. Initially, it was translated into English language using Google Translate. Afterwards, it was preprocessed by removing duplicates and missing values. After preprocessing, the final dataset consists of 5046 unique reports from 5046 patients with an average age of 53 years and 100% women. Furthermore, we used word-level NLP-based embedding techniques, term frequency-inverse document frequency (TF-IDF) and word2vec to extract semantic and syntactic information. We also compared the performance of ML, DL and large language models (LLMs) classifiers for BI-RADS category classification.ResultsThe final breast imaging-based radiological reports dataset contains 5046 unique reports. We compared K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Naive Bayes (NB), Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient-Boosting (GB), Extreme Gradient Boosting (XGB), Long Short-Term Memory (LSTM), Bidirectional Encoder Representations from Transformers (BERT) and Biomedical Generative Pre-trained Transformer (BioGPT) classifiers. It is observed that the BioGPT classifier with preprocessed data performed 6% better with a mean sensitivity of 0.60 (95% confidence interval (CI), 0.391-0.812) compared to the second best performing classifier BERT, which achieved mean sensitivity of 0.54 (95% CI, 0.477-0.607).ConclusionIn this work, we propose a curated and annotated benchmark dataset that can be used for BI-RADS and breast density category classification. We also provide baseline results of most ML, DL and LLMs models for BI-RADS classification that can be used as a starting point for future investigation. The main objective of this investigation is to provide a repository for the investigators who wish to enter the field to push the boundaries further.

Dual-energy CT late arterial phase iodine maps for the diagnosis of acute non-occlusive mesenteric ischemia.

To evaluate the diagnostic accuracy of dual-energy CT (DECT) iodine maps in comparison to conventional CT series for the assessment of non-occlusive mesenteric ischemia (NOMI). We evaluated data from 142 patients (72 men; 50.7%) who underwent DECT between 2018 and 2022, with surgically confirmed diagnosis of NOMI. One board-certified radiologist performed region of interest (ROI) measurements in bowel segments on late arterial (LA) and portal venous (PV) phase DECT iodine maps as well as LA conventional series, in both ischemic and non-ischemic bowel loops, using surgical reports as reference standard, and in a control group of 97 patients. Intra- and inter-reader agreement with a second board-certified radiologist was also evaluated. Receiver operating characteristic (ROC) curve analysis was performed to calculate the optimal threshold for discriminating ischemic from non-ischemic bowel segments. Subjective image rating of LA and PV iodine maps was performed. DECT-based iodine concentration (IC) measurements showed significant differences in LA phase iodine maps between ischemic (median:0.72; IQR 0.52-0.91mg/mL) and non-ischemic bowel loops (5.16; IQR 3.45-6.31mg/ml) (P <.0001). IC quantification on LA phase revealed an AUC of 0.966 for the assessment of acute bowel ischemia, significantly higher compared to both IC quantification based on PV phase (0.951) and attenuation values evaluated on LA conventional CT series (0.828). Excellent intra-observer and strong inter-observer agreements were observed for the quantification of iodine concentration. Conversely, weak inter-observer agreement was noted for conventional HU assessments. The optimal LA phase-based IC threshold for assessing bowel ischemia was 1.34mg/mL, yielding a sensitivity of 100% and specificity of 96.48%. Iodine maps based on LA phase significantly improve the diagnostic accuracy for the assessment of NOMI compared to conventional CT series and PV phase iodine maps.

Radiographic findings in dogs with 360 degrees gastric dilatation and volvulus.

Gastric dilatation and volvulus (GDV) is a life-threatening emergency that requires urgent intervention. Radiographic features associated with 360-GDV in dogs have not been investigated. The aim of this retrospective observational study is to describe radiographic features and clinical variables in dogs affected with 360-GDV and to report agreement rates between different radiologists. We also report the sensitivity and specificity of radiographs to diagnose 360-GDV in dogs. Confirmed 360-GDV cases were retrieved, and the radiographic findings were compared with dogs presenting with gastric dilatation (GD) and 180-GDV. Images were reviewed and graded by three blinded board-certified radiologists. A total of 16 dogs with confirmed 360-GDV were identified. The median age was 10 years old (2-13 years). The sensitivity for detection of 360-GDV ranged between 43.7% and 50%, and the specificity between 84.6% and 92.1%. Interobserver agreement on final diagnosis was substantial (Kappa=0.623; 0.487-0.760, 95% CI). The highest agreement rate was in cases of 180-GDV (87%), followed by the GD cases (72%) and 360-GDV (46%). Severe esophageal distension and absence of small intestinal dilation were the only radiographic features specifically associated with 360-GDV. A similar pyloric position was found between GD and 360-GDV. Additional radiographic variables that could help differentiate GD from 360-GDV include the degree of gastric distension and the peritoneal serosal contrast. Two cases with 360-GDV were misdiagnosed by the three radiologists as GD. In conclusion, radiographically, 360-GDV cases can reassemble GD and vice versa. Radiologists and clinicians should be aware of the low sensitivity of radiographs for the detection of 360-GDV.

Significantly improving zero-shot X-ray pathology classification via fine-tuning pre-trained image-text encoders

Deep neural networks are increasingly used in medical imaging for tasks such as pathological classification, but they face challenges due to the scarcity of high-quality, expert-labeled training data. Recent efforts have utilized pre-trained contrastive image-text models like CLIP, adapting them for medical use by fine-tuning the model with chest X-ray images and corresponding reports for zero-shot pathology classification, thus eliminating the need for pathology-specific annotations. However, most studies continue to use the same contrastive learning objectives as in the general domain, overlooking the multi-labeled nature of medical image-report pairs. In this paper, we propose a new fine-tuning strategy that includes positive-pair loss relaxation and random sentence sampling. We aim to improve the performance of zero-shot pathology classification without relying on external knowledge. Our method can be applied to any pre-trained contrastive image-text encoder and easily transferred to out-of-domain datasets without further training, as it does not use external data. Our approach consistently improves overall zero-shot pathology classification across four chest X-ray datasets and three pre-trained models, with an average macro AUROC increase of 4.3%. Additionally, our method outperforms the state-of-the-art and marginally surpasses board-certified radiologists in zero-shot classification for the five competition pathologies in the CheXpert dataset.

Computed tomography scan accuracy for the prediction of lobe and division of liver tumors by four board-certified radiologists.

(1) Evaluate the accuracy of computed tomography (CT) scans for localization of liver masses. (2) Assess the agreement between radiologists on localization. (3) Determine if location influences the accuracy of localization and histopathologic diagnosis. (4) Determine what lobar vasculature radiologists found most useful for localization. Retrospective. A total of 67 client-owned dogs with a total of 75 hepatic masses. Records were reviewed for relevant data. Localization for each hepatic mass was performed by four radiologists (JH, EH, ML, JF) independently. Overall accuracy of mass localization was 217/292 (74.3%) by lobe and 264/300 (88%) by division. Accuracy for the quadrate lobe (11/27, 40.7%) was lower (p < .05) than for the caudate process of the caudate lobe (19/24, 79.2%), left medial lobe (47/64, 73.4%) and left lateral lobe (95/101, 89.6%). Accuracy for the right lateral lobe (17/35, 48.6%) was lower (p < .05) lower than for the left lateral lobe (95/101, 89.6%). Accuracy of localization was 173/192 (90.1%) for masses located in the left division, 37/48 (77.1%) in the central division, and 53/60 (88.3%) for the right division. The agreement (kappa) between radiologists was good (0.61-0.8) to excellent (0.81-1) for division and moderate (0.41-0.6) to good for lobe localization. CT localization was more accurate for division than lobe localization of canine hepatic masses. Similarly, radiologists had a better agreement for division than lobe localization. This study supports CT as a useful modality for liver mass localization based on division. CT localization to specific lobes should be interpreted with some caution.

Evaluating the Performance and Bias of Natural Language Processing Tools in Labeling Chest Radiograph Reports.

Background Natural language processing (NLP) is commonly used to annotate radiology datasets for training deep learning (DL) models. However, the accuracy and potential biases of these NLP methods have not been thoroughly investigated, particularly across different demographic groups. Purpose To evaluate the accuracy and demographic bias of four NLP radiology report labeling tools on two chest radiograph datasets. Materials and Methods This retrospective study, performed between April 2022 and April 2024, evaluated chest radiograph report labeling using four NLP tools (CheXpert [rule-based], RadReportAnnotator [RRA; DL-based], OpenAI's GPT-4 [DL-based], cTAKES [hybrid]) on a subset of the Medical Information Mart for Intensive Care (MIMIC) chest radiograph dataset balanced for representation of age, sex, and race and ethnicity (n = 692) and the entire Indiana University (IU) chest radiograph dataset (n = 3665). Three board-certified radiologists annotated the chest radiograph reports for 14 thoracic disease labels. NLP tool performance was evaluated using several metrics, including accuracy and error rate. Bias was evaluated by comparing performance between demographic subgroups using the Pearson χ2 test. Results The IU dataset included 3665 patients (mean age, 49.7 years ± 17 [SD]; 1963 female), while the MIMIC dataset included 692 patients (mean age, 54.1 years ± 23.1; 357 female). All four NLP tools demonstrated high accuracy across findings in the IU and MIMIC datasets, as follows: CheXpert (92.6% [47 516 of 51 310], 90.2% [8742 of 9688]), RRA (82.9% [19 746 of 23 829], 92.2% [2870 of 3114]), GPT-4 (94.3% [45 586 of 48 342], 91.6% [6721 of 7336]), and cTAKES (84.7% [43 436 of 51 310], 88.7% [8597 of 9688]). RRA and cTAKES had higher accuracy (P < .001) on the MIMIC dataset, while CheXpert and GPT-4 had higher accuracy on the IU dataset. Differences (P < .001) in error rates were observed across age groups for all NLP tools except RRA on the MIMIC dataset, with the highest error rates for CheXpert, RRA, and cTAKES in patients older than 80 years (mean, 15.8% ± 5.0) and the highest error rate for GPT-4 in patients 60-80 years of age (8.3%). Conclusion Although commonly used NLP tools for chest radiograph report annotation are accurate when evaluating reports in aggregate, demographic subanalyses showed significant bias, with poorer performance in older patients. © RSNA, 2024 Supplemental material is available for this article. See also the editorial by Cai in this issue.

Diagnostic performance of CT/MRI LI-RADS v2018 in non-cirrhotic hepatitis C virus infection.

To evaluate the diagnostic performance of LI-RADS among patients with non-cirrhotic hepatitis C virus (HCV) infection. This retrospective, IRB-approved, single-center study included 66 observations from 43 adult patients (11 women, 32 men; median age 65 years). All patients received liver protocol CT or MRI from 2010 to 2023, had HCV, and did not have cirrhosis based on histopathology. Three board-certified abdominal radiologists blinded to histopathology and imaging follow-up assessed each observation for major features and final LI-RADS category, and inter-reader agreements with weighted kappa were calculated. The positive predictive value, sensitivity, specificity, and accuracy of in diagnosing HCC and overall malignancy was calculated. Of the 66 observations, 53 (80%) were malignant and 13 (20%) were benign. Positive predictive value for HCC was 0-0% for LR-1, 0-0% for LR-2, 0-33% for LR-3, 57-100% for LR-4, 98-100% for LR-5, 25-50% for LR-M, and 83-100% for LR-TIV. Positive predictive value for overall malignancy was 0-0% for LR-1, 0-0% for LR-2, 0-33% for LR-3, 57-100% for LR-4, 98-100% for LR-5, 100-100% for LR-M, and 100-100% for LR-TIV. For LR-5 in identifying HCC, sensitivity ranged from 74 to 90%, specificity from 94 to 100%, and accuracy from 80 to 91%. For the composite of LR-5, LR-M, or LR-TIV in identifying overall malignancy, sensitivity was 87-98%, specificity was 92-100%, and accuracy was 89-97%. The inter-reader agreement for major features varied from moderate to substantial, with substantial agreement for the final category. CT/MRI LI-RADS v2018 criteria can be applied to non-cirrhotic HCV patients with near-perfect specificity.

Performance of An Automated Algorithm in Large and Medium Vessel Occlusion Detection: A Real-World Experience.

Fast, accurate detection of large (LVO) and medium vessel occlusion (MeVO) is critical for triage and management of acute ischemic stroke. Multiple AIbased software products are commercially available. However, their strengths and limitations for detection of vessel occlusion in the context of expanding indications for mechanical thrombectomy are not entirely understood. We aimed to investigate the performance of a fully automated commercial detection algorithm to identify large and medium vessel occlusions in Code Stroke patients. We utilized a single-center, institutional, retrospective registry of all consecutive code stroke patients with CTA and automated processing using Viz.ai presenting at a comprehensive stroke center between March 2020 and February 2023. LVO was categorized as anterior LVO (aLVO), defined as occlusion of the intracranial internal cerebral artery or M1-middle cerebral artery (MCA), and posterior LVO (pLVO), defined as occlusion of the basilar artery or V4-vertebral artery. MeVO was defined as occlusion of the M2-MCA, A1/A2-anterior cerebral artery, or P1/P2-posterior cerebral artery. Reports from 12 board-certified radiologists were considered the gold standard. We analyzed the performance of the automated algorithm using STARD guidelines. Our primary outcome was accuracy of anterior LVO (aLVO) by the software. Secondary outcomes were accuracy of the software to detect three additional categories: all LVO (aLVO and pLVO), aLVO with M2-MCA, and aLVO with MeVO. Of 3,590 code stroke patients, 3,576 were technically sufficient for analysis by the automated software (median age 67 years; 51% female; 68% White), of which 616 (17.2%) had vessel occlusions. The respective sensitivity and specificity for all four pre-specified categories were: aLVO: 91% (87-94%), 93% (92-94%); all LVO: 73% (68-77%), 92% (91-93%); aLVO with M2-MCA:74% (70-78%), 93% (92-94%); aLVO with all MeVO: 65% (61-69%), and 93% (92-94%). The automated algorithm demonstrated high accuracy in identifying anterior LVO with lower performance for pLVO and MeVO. It is crucial for acute stroke teams to be aware of the discordance between automated algorithm results and true rates of LVO and MeVO for timely diagnosis and triage. LVO = large vessel occlusion; aLVO = anterior large vessel occlusion; pLVO = posterior large vessel occlusion; MeVO = medium vessel occlusion; EVT= endovascular thrombectomy; AI = artificial intelligence; ACA = anterior cerebral artery; PCA = posterior cerebral artery; BA = basilar artery; VA = vertebral artery.

Role of teleradiology in the interpretation of ultrasound images acquired in the emergency setting

BACKGROUND: Teleradiology has become an important tool in emergency medicine, particularly in the interpretation of emergency ultrasonography. In emergency situations, where time is essential, rapid diagnosis and treatment can mean the difference between life and death. Teleradiology is an innovative alternative to augment the staffing and fill the gaps of onsite radiology personnel in emergency departments or during off-hours. AIM: To assess the effectiveness/impact of teleradiology in emergency ultrasound interpretation. MATERIALS AND METHODS: A retrospective study was performed in a cohort of 33,616 patients from 86 hospitals across the USA between January and December 2022. The study involved radiological interpretations of 37,253 ultrasound images acquired in the emergency setting by American Board Certified Radiologists empaneled by a teleradiology service provider, headquartered in Bangalore, India. RESULTS: The proposed telehealth model provided timely and quality reporting of 37,253 scans of patients with a mean turnaround time of 35.71 min (95% confidence interval 35.50–35.91). CONCLUSION: This study demonstrates that a structured telesonography program with defined protocols for image capture, transmission, and clinical communication can allow for successful immediate reporting of ultrasound data in the emergency care setting.

Determinants of decision-making in biopsy of PI-RADS 3 transition zone lesions.

Cancer rates for Prostate Imaging-Reporting and Data System (PI-RADS) 3 lesions are low. We aimed to determine the clinical and magnetic resonance imaging (MRI) parameters that can provide risk stratification for PI-RADS 3 transition zone (TZ) lesions to guide decision for biopsy, which can improve the cost-effectiveness of resource utilisation. The MRI scans of all patients who underwent MRI-ultrasound fusion targeted biopsy from 1 May 2016 to 31 December 2022 were retrospectively assessed by two board-certified abdominal radiologists. The following data were collected and analysed serum prostate-specific antigen, Prostatic Health Index (PHI), prostate volume, histological results, lesion size, location, diffusion-weighted imaging (DWI) parameter scores and overall PI-RADS score. Two hundred and fourteen TZ lesions were included. Among 131 PI-RADS 3 lesions, those with marked restricted diffusion (DWI score ≥4), diameter ≥1 cm, prostrate-specific antigen density (PSAD) ≥0.11 and PHI ≥34 were more likely to contain clinically significant prostate cancer (csPCa; P = 0.04, 0.02, 0.049 and 0.05, respectively), with areas under the receiver operating characteristics curve of 0.9, 0.76, 0.84 and 0.80, respectively. Apical lesions were more likely to contain csPCa compared to midgland or basal lesions (P = 0.01). Clinical parameters (PSAD and PHI) and MRI features (lesion size, DWI score, lesion location) can be used to risk stratify PI-RADS 3 TZ lesions and guide decision for targeted biopsy.

Carotid artery assessment in dual-source photon-counting CT: impact of low-energy virtual monoenergetic imaging on image quality, vascular contrast and diagnostic assessability.

Preliminary dual-energy CT studies have shown that low-energy virtual monoenergetic (VMI) + reconstructions can provide superior image quality compared to standard 120kV CTA series. The purpose of this study is to evaluate the impact of low-energy VMI reconstructions on quantitative and qualitative image quality, vascular contrast, and diagnostic assessability of the carotid artery in patients undergoing photon-counting CTA examinations. A total of 122 patients (67 male) who had undergone dual-source photon-counting CTA scans of the carotid artery were retrospectively analyzed in this study. Standard 120kV CT images and low-keV VMI series from 40 to 100keV with an interval of 15keV were reconstructed. Quantitative analyses included the evaluation of vascular CT numbers, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). CT number measurements were performed in the common, external, and internal carotid arteries. Qualitative analyses were performed by three board-certified radiologists independently using five-point scales to evaluate image quality, vascular contrast, and diagnostic assessability of the carotid artery. Mean attenuation, CNR and SNR values were highest in 40keV VMI reconstructions (HU, 1362.32 ± 457.81; CNR, 33.19 ± 12.86; SNR, 34.37 ± 12.89) followed by 55-keV VMI reconstructions (HU, 736.94 ± 150.09; CNR, 24.49 ± 7.11; SNR, 26.25 ± 7.34); all three mean values at these keV levels were significantly higher compared with the remaining VMI series and standard 120kV CT series (HU, 154.43 ± 23.69; CNR, 16.34 ± 5.47; SNR, 24.44 ± 7.14) (p < 0.0001). The qualitative analysis showed the highest rating scores for 55keV VMI reconstructions followed by 40keV and 70keV VMI series with a significant difference compared to standard 120kV CT images series regarding image quality, vascular contrast, and diagnostic assessability of the carotid artery (all comparisons, p < 0.01). Low-keV VMI reconstructions at a level of 40-55keV significantly improve image quality, vascular contrast, and the diagnostic assessability of the carotid artery compared with standard CT series in photon-counting CTA.

The Diagnostic Performance of Large Language Models and General Radiologists in Thoracic Radiology Cases: A Comparative Study.

To investigate and compare the diagnostic performance of 10 different large language models (LLMs) and 2 board-certified general radiologists in thoracic radiology cases published by The Society of Thoracic Radiology. We collected publicly available 124 "Case of the Month" from the Society of Thoracic Radiology website between March 2012 and December 2023. Medical history and imaging findings were input into LLMs for diagnosis and differential diagnosis, while radiologists independently visually provided their assessments. Cases were categorized anatomically (parenchyma, airways, mediastinum-pleura-chest wall, and vascular) and further classified as specific or nonspecific for radiologic diagnosis. Diagnostic accuracy and differential diagnosis scores (DDxScore) were analyzed using the χ2, Kruskal-Wallis, Wilcoxon, McNemar, and Mann-Whitney U tests. Among the 124 cases, Claude 3 Opus showed the highest diagnostic accuracy (70.29%), followed by ChatGPT 4/Google Gemini 1.5 Pro (59.75%), Meta Llama 3 70b (57.3%), ChatGPT 3.5 (53.2%), outperforming radiologists (52.4% and 41.1%) and other LLMs (P<0.05). Claude 3 Opus DDxScore was significantly better than other LLMs and radiologists, except ChatGPT 3.5 (P<0.05). All LLMs and radiologists showed greater accuracy in specific cases (P<0.05), with no DDxScore difference for Perplexity and Google Bard based on specificity (P>0.05). There were no significant differences between LLMs and radiologists in the diagnostic accuracy of anatomic subgroups (P>0.05), except for Meta Llama 3 70b in the vascular cases (P=0.040). Claude 3 Opus outperformed other LLMs and radiologists in text-based thoracic radiology cases. LLMs hold great promise for clinical decision systems under proper medical supervision.

Revealing the surprising: Computed tomography-guided exploration of paranasal sinuses incidental findings

BackgroundIncidental findings (IFs) are unintentional discoveries that are unrelated to the original imaging goal. Imaging exams of persons with suspected intracranial disorders may reveal IFs in the paranasal sinuses (PNS). The current investigation aims to determine the frequency and features of unexpected discoveries in the PNS in persons who have had brain computed tomography (CT) scans for suspected intracranial abnormalities. MethodsBetween December 2022 and February 2023, 100 patients who met the inclusion criteria for clinically suspected intracranial disorders had CT brain scans as part of this retrospective cross-sectional study. Two board-certified radiologists with at least three years of experience assessed the CT scans independently. The study investigated the incidence and proportion of paranasal sinuses incidental findings (PNS IFs) observed during brain CT scans to detect brain abnormalities. ResultsThe study discovered that 27% of patients had IFs, with the retired population (aged 61 and up) having the highest prevalence, particularly among men. Acute sinusitis was the most commonly diagnosed incidental finding (IF), accounting for 15% of total cases. In addition, we found polyps, retention cysts, chronic sinusitis, mucoceles, and fungal infections. Left-sided maxillary sinus abnormalities outnumbered right-sided ones. Bilateral involvement was unusual. ConclusionThese findings emphasize the importance of addressing IFs in the PNS when treating patients, as they may necessitate further inspection or intervention. These results have the potential to help establish strategies for treating patients with incidental paranasal sinus findings, ultimately improving patient care.

Screening Patient Misidentification Errors Using a Deep Learning Model of Chest Radiography: A Seven Reader Study.

We aimed to evaluate the ability of deep learning (DL) models to identify patients from a paired chest radiograph (CXR) and compare their performance with that of human experts. In this retrospective study, patient identification DL models were developed using 240,004 CXRs. The models were validated using multiple datasets, namely, internal validation, CheXpert, and Chest ImaGenome (CIG), which include different populations. Model performance was analyzed in terms of disease change status. The performance of the models to identify patients from paired CXRs was compared with three junior radiology residents (group I), two senior radiology residents (group II), and two board-certified expert radiologists (group III). For the reader study, 240 patients (age, 56.617 ± 13.690years, 113 females, 160 same pairs) were evaluated. A one-sided non-inferiority test was performed with a one-sided margin of 0.05. SimChest, our similarity-based DL model, demonstrated the best patient identification performance across multiple datasets, regardless of disease change status (internal validation [area under the receiver operating characteristic curve range: 0.992-0.999], CheXpert [0.933-0.948], and CIG [0.949-0.951]). The radiologists identified patients from the paired CXRs with a mean accuracy of 0.900 (95% confidence interval: 0.852-0.948), with performance increasing with experience (mean accuracy:group I [0.874], group II [0.904], group III [0.935], and SimChest [0.904]). SimChest achieved non-inferior performance compared to the radiologists (P for non-inferiority: 0.015). The findings of this diagnostic study indicate that DL models can screen for patient misidentification using a pair of CXRs non-inferiorly to human experts.

Verification of image quality improvement by deep learning reconstruction to 1.5T MRI in T2-weighted images of the prostate gland.

This study aimed to evaluate whether the image quality of 1.5T magnetic resonance imaging (MRI) of the prostate is equal to or higher than that of 3T MRI by applying deep learning reconstruction (DLR). To objectively analyze the images from the 13 healthy volunteers, we measured the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) of the images obtained by the 1.5T scanner with and without DLR, as well as for images obtained by the 3T scanner. In the subjective, T2W images of the prostate were visually evaluated by two board-certified radiologists. The SNRs and CNRs in 1.5T images with DLR were higher than that in 3T images. Subjective image scores were better for 1.5T images with DLR than 3T images. The use of the DLR technique in 1.5T MRI substantially improved the SNR and image quality of T2W images of the prostate gland, as compared to 3T MRI.

Prediction of Clinically Significant Prostate Cancer Using Multiparametric MRI, Biparametric MRI, and Clinical Parameters

Introduction: Multiparametric MRI (mpMRI) is gold standard for the primary diagnostic work-up of clinically significant prostate cancer (csPCa). The aim of this study was to assess the benefit of the perfusion sequence and the non-inferiority of an MRI without contrast administration (bpMRI) compared to mpMRI while taking clinical parameters into account. Methods: In this retrospective, non-interventional study we examined MRI data from 355 biopsy-naïve patients, performed on a 3T MRI system, evaluated by a board-certified radiologist with over 10 years of experience with subsequent mpMRI-TRUS fusion biopsy. Discussion: Only 16/355 (4.5%) patients benefited from dynamic contrast enhanced. In only 3/355 (0.8%) patients, csPCa would have been missed in bpMRI. BpMRI provided sensitivity and specificity (81.4%; 79.4%) comparable to mpMRI (75.2%; 81.8%). Additionally, bpMRI and mpMRI were independent predictors for the presence of csPCa, individually (OR: 15.36; p < 0.001 vs. 12.15; p = 0.006) and after accounting for established influencing factors (OR: 12.81; p < 0.001 vs. 6.50; p = 0.012). When clinical parameters were considered, a more balanced diagnostic performance between sensitivity and specificity was found for mpMRI and bpMRI. Overall, PSA density showed the highest diagnostic performance (area under the curve = 0.81) for the detection of csPCa. Conclusion: The premise of the study was confirmed. Therefore, bpMRI should be adopted as soon as existing limitations have been lifted by prospective multi-reader studies.

Diagnostic Performance of a Deep Learning-Powered Application for Aortic Dissection Triage Prioritization and Classification.

This multicenter retrospective study evaluated the diagnostic performance of a deep learning (DL)-based application for detecting, classifying, and highlighting suspected aortic dissections (ADs) on chest and thoraco-abdominal CT angiography (CTA) scans. CTA scans from over 200 U.S. and European cities acquired on 52 scanner models from six manufacturers were retrospectively collected and processed by CINA-CHEST (AD) (Avicenna.AI, La Ciotat, France) device. The diagnostic performance of the device was compared with the ground truth established by the majority agreement of three U.S. board-certified radiologists. Furthermore, the DL algorithm's time to notification was evaluated to demonstrate clinical effectiveness. The study included 1303 CTAs (mean age 58.8 ± 16.4 years old, 46.7% male, 10.5% positive). The device demonstrated a sensitivity of 94.2% [95% CI: 88.8-97.5%] and a specificity of 97.3% [95% CI: 96.2-98.1%]. The application classified positive cases by the AD type with an accuracy of 99.5% [95% CI: 98.9-99.8%] for type A and 97.5 [95% CI: 96.4-98.3%] for type B. The application did not miss any type A cases. The device flagged 32 cases incorrectly, primarily due to acquisition artefacts and aortic pathologies mimicking AD. The mean time to process and notify of potential AD cases was 27.9 ± 8.7 s. This deep learning-based application demonstrated a strong performance in detecting and classifying aortic dissection cases, potentially enabling faster triage of these urgent cases in clinical settings.

Invasive mould infection in immunocompromised children: Pulmonary computed tomography findings from two multicentre paediatric cohorts

BackgroundInvasive fungal disease (IFD) can cause significant morbidity in immunocompromised children. The lung is the most common site for IFD caused by moulds. Timely diagnosis is key to facilitate initiation of appropriate treatment. As clinical manifestations are non-specific, initial diagnosis relies on computed tomography (CT) imaging, yet there are limited recent data documenting imaging changes in paediatric pulmonary IFD. In this study, CT images of children with proven/probable pulmonary IFD from two multi-centre cohorts were assessed to document typical changes, including a comparison between causative pathogens, and to measure reliability of radiological reporting in this context. MethodsPulmonary IFD cases diagnosed between 2003–2013 and 2017–2020 were identified from two respective Australian multi-centre cohort studies of immunocompromised children. Proven/probable IFD was defined according to the EORTC/MSG definitions. Initial CT images at the time of IFD diagnosis were independently reviewed by two board-certified paediatric radiologists. Presence of EORTC/MSG-defined and other changes was documented. Inter-rater reliability was assessed. Impact of specific imaging features on outcome at 6 months was assessed using multivariate regression. ResultsA total of 99 cases were included (65 aspergillosis, 19 non-aspergillus moulds (NAM) and 15 histopathological-mould cases). Median age was 9.6 years (IQR 5.1–13.6); 89.9 % had an underlying haematological malignancy. Nodules (85.4 %), the halo sign (49.5 %) and wedge shaped/segmental consolidation (35.9 %) were the most frequent EORTC/MSG-defined features. Cavitation and air-crescent sign were uncommon. Agreement between radiologists was fair to moderate for most common imaging features. No specific radiological finding differentiated aspergillosis from NAM infection. The presence of mass on initial CT and NAM infection were independently associated with poor outcome. ConclusionsIn proven/probable pulmonary IFD in immunocompromised children, nodules with or without halo sign are the most frequent finding. Agreement between radiologists was variable for common imaging features, and specific findings did not correlate with the causative pathogen. Microbiological confirmation remains vital in children with suspected pulmonary IFD.

A vision-language foundation model for the generation of realistic chest X-ray images.

The paucity of high-quality medical imaging datasets could be mitigated by machine learning models that generate compositionally diverse images that faithfully represent medical concepts and pathologies. However, large vision-language models are trained on natural images, and the diversity distribution of the generated images substantially differs from that of medical images. Moreover, medical language involves specific and semantically rich vocabulary. Here we describe a domain-adaptation strategy for large vision-language models that overcomes distributional shifts. Specifically, by leveraging publicly available datasets of chest X-ray images and the corresponding radiology reports, we adapted a latent diffusion model pre-trained on pairs of natural images and text descriptors to generate diverse and visually plausible synthetic chest X-ray images (as confirmed by board-certified radiologists) whose appearance can be controlled with free-form medical text prompts. The domain-adaptation strategy for the text-conditioned synthesis of medical images can be used to augment training datasets and is a viable alternative to the sharing of real medical images for model training and fine-tuning.