Junior Radiologists Research Articles

Assessing the Performance of ChatGPT and Bard/Gemini Against Radiologists for PI-RADS Classification Based on Prostate Multiparametric MRI Text Reports.

Large language models (LLMs) have shown potential for clinical applications. This study assesses their ability to assign PI-RADS categories based on clinical text reports. One hundred consecutive biopsy-naïve patients' multiparametric prostate MRI reports were independently classified by two uroradiologists, GPT-3.5, GPT-4, Bard, and Gemini. Original report classifications were considered definitive. Out of 100 MRIs, 52 were originally reported as PI-RADS 1-2, 9 PI-RADS 3, 19 PI-RADS 4, and 20 PI-RADS 5. Radiologists demonstrated 95% and 90% accuracy, while GPT-3.5 and Bard both achieved 67%. Accuracy of the updated versions of LLMs increased to 83% (GTP-4) and 79% (Gemini), respectively. In low suspicion studies (PI-RADS 1-2), Bard and Gemini (F1: 0.94, 0.98, respectively) outperformed GPT-3.5 and GTP-4 (F1:0.77, 0.94, respectively), whereas for high probability MRIs (PI-RADS 4-5), GPT-3.5 and GTP-4 (F1: 0.95, 0.98, respectively) outperformed Bard and Gemini (F1: 0.71, 0.87, respectively). Bard assigned a non-existent PI-RADS 6 "hallucination" for two patients. Inter-reader agreements (Κ) between the original reports and the senior radiologist, junior radiologist, GPT-3.5, GTP-4, BARD, and Gemini were 0.93, 0.84, 0.65, 0.86, 0.57, and 0.81, respectively. Radiologists demonstrated high accuracy in PI-RADS classification based on text reports, while GPT-3.5 and Bard exhibited poor performance. GTP-4 and Gemini demonstrated improved performance compared to their predecessors. This study highlights the limitations of LLMs in accurately classifying PI-RADS categories from clinical text reports. While the performance of LLMs has improved with newer versions, caution is warranted before integrating such technologies into clinical practice.

Machine learning and deep learning models for preoperative detection of lymph node metastasis in colorectal cancer: a systematic review and meta-analysis.

To evaluate the diagnostic performance of Machine Learning (ML) and Deep Learning (DL) models for predicting preoperative Lymph Node Metastasis (LNM) in Colorectal Cancer (CRC) patients. A systematic review and meta-analysis were conducted following PRISMA-DTA and AMSTAR-2 guidelines. We searched PubMed, Web of Science, Embase, and Cochrane Library databases until February 16, 2024. Study quality and risk of bias were assessed using the QUADAS-2 tool. Data were analyzed using STATA v18, applying random-effects models to all analyses. Twelve studies involving 8321 patients were included, with most published in 2021-2024 (9/12). The pooled AUC of ML models for predicting LNM in CRC patients was 0.87 (95% CI: 0.82-0.91, I2:86.17) with a sensitivity of 78% (95% CI: 69-87%) and a specificity of 77% (95% CI: 64%-90%). In addition, when assessing the AUC reported by radiologists, both junior and senior radiologists had similar performance, significantly lower than the ML models. (P < 0.001). Subgroup analysis revealed higher AUCs in prospective studies (0.95, 95% CI: 0.87-1) compared to retrospective studies (0.85, 95% CI: 0.81-0.89) (P = 0.03). Studies without external validation exhibited significantly higher AUCs than those with external validation (P < 0.01). While there was no significant difference in AUC and sensitivity between the T1-T2 and T2-T4 stages, specificity was significantly higher in the T2-T4 stages than the low stages of T1 and T2 (95%, 95% CI: 92-98% vs. 61%, 95% CI: 44-78%; P < 0.01). ML models demonstrate strong potential for preoperative LNM staging and treatment planning in CRC, potentially reducing the need for additional surgeries and related health and financial burdens. Further prospective multicenter studies, with standardized reporting of algorithms, modality parameters, and LNM staging, are needed to validate these findings.

Automatic segmentation and visualization of cortical and marrow bone in mandibular condyle on CBCT: a preliminary exploration of clinical application.

To develop a deep learning-based automatic segmentation method for cortex and marrow in mandibular condyle on cone-beam computed tomography (CBCT) images and explore its clinical application. 825 condyles of 490 CBCT images from 3 centers of Stomatology hospitalaffliated to Zhejiang University School of Medicinewere collected. A deep learning model was developed for simultaneous segmentation of cortex and marrow in mandibular condyle. It included a region of interest extraction network and a segmentation network based on 3D U-net, with modifications made to improve the segmentation boundaries. To evaluate its clinical potential, the model's segmentation efficiency and accuracy were compared with those of both junior and senior oral and maxillofacial radiologists. Additionally, the model's ability to assist junior radiologists in diagnosis through visualization and quantitative analysis of the generated 3D model was also assessed. The Dice similarity coefficient of the deep learning model was 0.901 (cortex), 0.969 (marrow), and 0.982 (entire condyle). Hausdorff distance was 0.755mm (cortex), 0.826mm (marrow), and 0.760mm (entire condyle). The model outperformed radiologists across all segmentation metrics, completing the task in merely 15.06s. With the assistance of visualization and quantitative analysis generated from the model's segmentation, the diagnostic accuracy of junior radiologists significantly improved. The proposed deep learning-based model achieved accurate and efficient segmentation for mandibular condylar cortex and marrow. It possessed capability to generate precise 3D models, facilitating visual quantitative measurement and aiding in the diagnosis of condylar bony changes. This model holds potential for clinical applications in orthognathic surgery, orthodontic treatment, and other TMJ-related interventions.

Integrated Deep Learning Model for the Detection, Segmentation, and Morphologic Analysis of Intracranial Aneurysms Using CT Angiography.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop a deep learning model for the morphologic measurement of unruptured intracranial aneurysms (UIAs) based on CT angiography (CTA) data and validate its performance using a multicenter dataset. Materials and Methods In this retrospective study, patients with CTA examinations, including those with and without UIAs, in a tertiary referral hospital from February 2018 to February 2021 were included as the training dataset. Patients with UIAs who underwent CTA at multiple centers between April 2021 to December 2022 were included as the multicenter external testing set. An integrated deep-learning (IDL) model was developed for UIA detection, segmentation and morphologic measurement using an nnU-net algorithm. Model performance was evaluated using the Dice similarity coefficient (DSC) and intraclass correlation coefficient (ICC), with measurements by senior radiologists serving as the reference standard. The ability of the IDL model to improve performance of junior radiologists in measuring morphologic UIA features was assessed. Results The study included 1182 patients with UIAs and 578 controls without UIAs as the training dataset (55 years [IQR, 47-62], 1,012 [57.5%] females) and 535 patients with UIAs as the multicenter external testing set (57 years [IQR, 50-63], 353 [66.0%] females). The IDL model achieved 97% accuracy in detecting UIAs and achieved a DSC of 0.90 (95%CI, 0.88-0.92) for UIA segmentation. Model-based morphologic measurements showed good agreement with reference standard measurements (all ICCs > 0.85). Within the multicenter external testing set, the IDL model also showed agreement with reference standard measurements (all ICCs > 0.80). Junior radiologists assisted by the IDL model showed significantly improved performance in measuring UIA size (ICC improved from 0.88 [0.80-0.92] to 0.96 [0.92-0.97], P < .001). Conclusion The developed integrated deep learning model using CTA data showed good performance in UIA detection, segmentation and morphologic measurement and may be used to assist less experienced radiologists in morphologic analysis of UIAs. ©RSNA, 2024.

Development of fully automated models for staging liver fibrosis using non-contrast MRI and artificial intelligence: a retrospective multicenter study

Accurate staging of liver fibrosis (LF) is essential for clinical management in chronic liver disease. While non-contrast MRI (NC-MRI) yields valuable information for liver assessment, its effectiveness in predicting LF remains underexplored. This study aimed to develop and validate artificial intelligence (AI)-powered models utilizing NC-MRI for staging LF. A total of 1726 patients from Shengjing Hospital of China Medical University, registered between October 2003 and October 2022, were retrospectively collected, and divided into development (n=1208) and internal test (n=518) cohorts. An external test cohort consisting of 337 individuals from six centers, registered between June 2015 and November 2022, were also included. All participants underwent NC-MRI (T1-weighted imaging, T1WI; and T2-fat-suppressed imaging, T2FS) and liver biopsies. Two classification models (CMs), named T1 and T2FS, were trained on respective image types using 3D contextual transformer networks and evaluated on both test cohorts. Additionally, three CMs-Clinic, Image, and Fusion-were developed using clinical features, T1 and T2FS scores, and their integration via logistic regression. Classification effectiveness of CMs was assessed using the area under the receiver operating characteristic curve (AUC). A comparison was conducted between the optimal models (OMs) with highest AUC and other methods (transient elastography, five serum biomarkers, and six radiologists). Fusion models (i.e., OM) yielded the highest AUC among the CMs, achieving AUCs of 0.810 for significant fibrosis, 0.881 for advanced fibrosis, and 0.918 for cirrhosis in the internal test cohort, and 0.808, 0.868, and 0.925, respectively, in the external test cohort. The OMs demonstrated superior performance in AUC, significantly surpassing transient elastography (only for staging≥F2 and ≥ F3 grades), serum biomarkers, and three junior radiologists for staging LF. Radiologists, with the aid of the OMs, can achieve a higher AUC in LF assessment. AI-powered models utilizing NC-MRI, including T1WI and T2FS, accurately stage LF. National Natural Science Foundation of China (No. 82071885); General Program of the Liaoning Provincial Department of Education (LJKMZ20221160); Liaoning Province Science and Technology Joint Plan (2023JH2/101700127); the Leading Young Talent Program of Xingliao Yingcai in Liaoning Province (XLYC2203037).

Bridging Human and Machine Intelligence: Reverse-Engineering Radiologist Intentions for Clinical Trust and Adoption

In the rapidly evolving landscape of medical imaging, the integration of artificial intelligence (AI) with clinical expertise offers unprecedented opportunities to enhance diagnostic precision and accuracy. Yet, the "black box" nature of AI models often limits their integration into clinical practice, where transparency and interpretability are important. This paper presents a novel system leveraging the Large Multimodal Model (LMM) to bridge the gap between AI predictions and the cognitive processes of radiologists. This system consists of two core modules, Temporally Grounded Intention Detection (TGID) and Region Extraction (RE). The TGID module predicts the radiologist's intentions by analyzing eye gaze fixation heatmap videos and corresponding radiology reports. Additionally, the RE module extracts regions of interest that align with these intentions, mirroring the radiologist’s diagnostic focus. This approach introduces a new task, radiologist intention detection, and is the first application of Dense Video Captioning (DVC) in the medical domain. By making AI systems more interpretable and aligned with radiologist’s cognitive processes, this proposed system aims to enhance trust, improve diagnostic accuracy, and support medical education. Additionally, it holds the potential for automated error correction, guiding junior radiologists, and fostering more effective training and feedback mechanisms. This work sets a precedent for future research in AI-driven healthcare, offering a pathway towards transparent, trustworthy, and human-centered AI systems. We evaluated this model using NLG(Natural Language Generation), time-related, and vision-based metrics, demonstrating superior performance in generating temporally grounded intentions on REFLACX and EGD-CXR datasets. This model also demonstrated strong predictive accuracy in overlap scores for medical abnormalities and effective region extraction with high IoU(Intersection over Union), especially in complex cases like cardiomegaly and edema. These results highlight the system's potential to enhance diagnostic accuracy and support continuous learning in radiology. We are also releasing the source code for our project, available here.

Automatic tumor segmentation and lymph node metastasis prediction in papillary thyroid carcinoma using ultrasound keyframes.

Accurate preoperative prediction of cervical lymph node metastasis (LNM) for papillary thyroid carcinoma (PTC) patients is essential for disease staging and individualized treatment planning, which can improve prognosis and facilitate better management. To establish a fully automated deep learning-enabled model (FADLM) for automated tumor segmentation and cervical LNM prediction in PTC using ultrasound (US) video keyframes. The bicentral study retrospective enrolled 518 PTC patients, who were then randomly divided into the training (Hospital 1, n=340), internal test (Hospital 1, n=83), and external test cohorts (Hospital 2, n=95). The FADLM integrated mask region-based convolutional neural network (Mask R-CNN) for automatic thyroid primary tumor segmentation and ResNet34 with Bayes strategy for cervical LNM diagnosis. A radiomics model (RM) using the same automated segmentation method, a traditional radiomics model (TRM) using manual segmentation, and a clinical-semantic model (CSM) were developed for comparison. The dice similarity coefficient (DSC) was used to evaluate segmentation performance. The prediction performance of the models was validated in terms of discrimination and clinical utility with the area under the receiver operator characteristic curve (AUC), heatmap analysis, and decision curve analysis (DCA). The comparison of the predictive performance among different models was conducted by DeLong test. The performances of two radiologists compared with FADLM and the diagnostic augmentation with FADLM's assistance were analyzed in terms of accuracy, sensitivity and specificity using McNemar's x2 test. The p-value less than 0.05 was defined as a statistically significant difference. The Benjamini-Hochberg procedure was applied for multiple comparisons to deal with Type I error. The FADLM yielded promising segmentation results in training (DSC: 0.88±0.23), internal test (DSC: 0.88±0.23), and external test cohorts (DSC: 0.85±0.24). The AUCs of FADLM for cervical LNM prediction were 0.78 (95% CI: 0.73, 0.83), 0.83 (95% CI: 0.74, 0.92), and 0.83 (95% CI: 0.75, 0.92), respectively. It all significantly outperformed the RM (AUCs: 0.78vs. 0.72; 0.83vs. 0.65; 0.83vs. 0.68, all adjusted p-values<0.05) and CSM (AUCs: 0.78vs. 0.71; 0.83vs. 0.62; 0.83vs. 0.68, all adjusted p-values<0.05) across the three cohorts. The RM offered similar performance to that of TRM (AUCs: 0.61vs. 0.63, adjusted p-value =0.60) while significantly reducing the segmentation time (3.3±3.8 vs. 14.1±4.2 s, p-value <0.001). Under the assistance of FADLM, the accuracies of junior and senior radiologists were improved by 18% and 15% (all adjusted p-values<0.05) and the sensitivities by 25% and 21% (all adjusted p-values<0.05) in the external test cohort. The FADLM with elaborately designed automated strategy using US video keyframes holds good potential to provide an efficient and consistent prediction of cervical LNM in PTC. The FADLM displays superior performance to RM, CSM, and radiologists with promising efficacy.

Deep learning-based segmentation of kidneys and renal cysts on T2-weighted MRI from patients with autosomal dominant polycystic kidney disease

BackgroundOur aim was to train and test a deep learning-based algorithm for automatically segmenting kidneys and renal cysts in patients with autosomal dominant polycystic kidney disease (ADPKD).MethodsWe retrospectively selected all ADPKD patients who underwent renal MRI with coronal T2-weighted imaging at our institution from 2008 to 2022. The 20 most recent examinations constituted the test dataset, to mimic pseudoprospective enrolment. The remaining ones constituted the training dataset to which eight normal renal MRIs were added. Kidneys and cysts ground truth segmentations were performed on coronal T2-weighted images by a junior radiologist supervised by an experienced radiologist. Kidneys and cysts of the 20 test MRIs were segmented by the algorithm and three independent human raters. Segmentations were compared using overlap metrics. The total kidney volume (TKV), total cystic volume (TCV), and cystic index (TCV divided by TKV) were compared using Bland–Altman analysis.ResultsWe included 164 ADPKD patients. Dice similarity coefficients ranged from 85.9% to 87.4% between the algorithms and the raters’ segmentations and from 84.2% to 86.2% across raters’ segmentations. For TCV assessment, the biases ± standard deviations (SD) were 3–19 ± 137–151 mL between the algorithm and the raters, and 22–45 ± 49–57 mL across raters. The algorithm underestimated TKV and TCV in two outliers with TCV > 2800 mL. For cystic index assessment, the biases ± SD were 2.5–6.9% ± 6.7–8.3% between the algorithm and the raters, and 2.1–9.4 ± 7.4–11.6% across raters.ConclusionThe algorithm’s performance fell within the range of inter-rater variability, but large TKV and TCV were underestimated.Relevance statementAccurate automated segmentation of the renal cysts will enable the large-scale evaluation of the prognostic value of TCV and cystic index in ADPKD patients. If these biomarkers are prognostic, then automated segmentation will facilitate their use in daily routine.Key PointsCystic volume is an emerging biomarker in ADPKD.The algorithm’s performance in segmenting kidneys and cysts fell within interrater variability.The segmentation of very large cysts, under-represented in the training dataset, needs improvement.Graphical

Synthetic double inversion recovery imaging for rectal cancer T staging evaluation: imaging quality and added value to T2-weighted imaging

ObjectiveTo assess the image quality of synthetic double inversion recovery (SyDIR) imaging and enhance the value of T2-weighted imaging (T2WI) in evaluating T stage for rectal cancer patients.MethodsA total of 112 pathologically confirmed rectal cancer patients were retrospectively selected after undergoing MRI, including synthetic MRI. The image quality of T2WI and SyDIR imaging was compared based on signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), overall picture quality, presence of motion artifacts, lesion edge sharpness, and conspicuity. The concordance between MRI and pathological staging results, using T2WI alone and the combination of T2WI and SyDIR for junior and senior radiologists, was assessed using the Kappa test. The area under the receiver operating characteristic curve (AUC) was used to assess the diagnostic efficacy of extramural infiltration in rectal cancer patients.ResultsNo significant differences in imaging quality were observed between conventional T2WI and SyDIR (p = 0.07–0.53). The combination of T2WI and SyDIR notably improved the staging concordance between MRI and pathology for both junior (kappa value from 0.547 to 0.780) and senior radiologists (kappa value from 0.738 to 0.834). In addition, the integration of T2WI and SyDIR increased the AUC for diagnosing extramural infiltration for both junior (from 0.842 to 0.918) and senior radiologists (from 0.917 to 0.938).ConclusionThe combination of T2WI and SyDIR increased the consistency of T staging between MRI and pathology, as well as the diagnostic performance of extramural infiltration, which would benefit treatment selection.Critical relevance statementSyDIR sequence provides additional diagnostic value for T2WI in the T staging of rectal cancer, improving the agreement of T staging between MRI and pathology, as well as the diagnostic performance of extramural infiltration.Key PointsSynthetic double inversion recovery (SyDIR) and T2WI have comparable image quality.SyDIR provides rectal cancer anatomical features for extramural infiltration detections.The combination of T2WI and SyDIR improves the accuracy of T staging in rectal cancer.Graphical

Assessing the feasibility of ChatGPT-4o and Claude 3-Opus in thyroid nodule classification based on ultrasound images.

Large language models (LLMs) are pivotal in artificial intelligence, demonstrating advanced capabilities in natural language understanding and multimodal interactions, with significant potential in medical applications. This study explores the feasibility and efficacy of LLMs, specifically ChatGPT-4o and Claude 3-Opus, in classifying thyroid nodules using ultrasound images. This study included 112 patients with a total of 116 thyroid nodules, comprising 75 benign and 41 malignant cases. Ultrasound images of these nodules were analyzed using ChatGPT-4o and Claude 3-Opus to diagnose the benign or malignant nature of the nodules. An independent evaluation by a junior radiologist was also conducted. Diagnostic performance was assessed using Cohen's Kappa and receiver operating characteristic (ROC) curve analysis, referencing pathological diagnoses. ChatGPT-4o demonstrated poor agreement with pathological results (Kappa = 0.116), while Claude 3-Opus showed even lower agreement (Kappa = 0.034). The junior radiologist exhibited moderate agreement (Kappa = 0.450). ChatGPT-4o achieved an area under the ROC curve (AUC) of 57.0% (95% CI: 48.6-65.5%), slightly outperforming Claude 3-Opus (AUC of 52.0%, 95% CI: 43.2-60.9%). In contrast, the junior radiologist achieved a significantly higher AUC of 72.4% (95% CI: 63.7-81.1%). The unnecessary biopsy rates were 41.4% for ChatGPT-4o, 43.1% for Claude 3-Opus, and 12.1% for the junior radiologist. While LLMs such as ChatGPT-4o and Claude 3-Opus show promise for future applications in medical imaging, their current use in clinical diagnostics should be approached cautiously due to their limited accuracy.

Using Machine Learning on MRI Radiomics to Diagnose Parotid Tumours Before Comparing Performance with Radiologists: A Pilot Study.

The parotid glands are the largest of the major salivary glands. They can harbour both benign and malignant tumours. Preoperative work-up relies on MR images and fine needle aspiration biopsy, but these diagnostic tools have low sensitivity and specificity, often leading to surgery for diagnostic purposes. The aim of this paper is (1) to develop a machine learning algorithm based on MR images characteristics to automatically classify parotid gland tumours and (2) compare its results with the diagnoses of junior and senior radiologists in order to evaluate its utility in routine practice. While automatic algorithms applied to parotid tumours classification have been developed in the past, we believe that our study is one of the first to leverage four different MRI sequences and propose a comparison with clinicians. In this study, we leverage data coming from a cohort of 134 patients treated for benign or malignant parotid tumours. Using radiomics extracted from the MR images of the gland, we train a random forest and a logistic regression to predict the corresponding histopathological subtypes. On the test set, the best results are given by the random forest: we obtain a 0.720 accuracy, a 0.860 specificity, and a 0.720 sensitivity over all histopathological subtypes, with an average AUC of 0.838. When considering the discrimination between benign and malignant tumours, the algorithm results in a 0.760 accuracy and a 0.769 AUC, both on test set. Moreover, the clinical experiment shows that our model helps to improve diagnostic abilities of junior radiologists as their sensitivity and accuracy raised by 6 % when using our proposed method. This algorithm may be useful for training of physicians. Radiomics with a machine learning algorithm may help improve discrimination between benign and malignant parotid tumours, decreasing the need for diagnostic surgery. Further studies are warranted to validate our algorithm for routine use.

Machine learning model based on dynamic contrast-enhanced ultrasound assisting LI-RADS diagnosis of HCC: A multicenter diagnostic study

BackgroundTo enhance the accuracy of hepatocellular carcinoma (HCC) diagnosis using contrast-enhanced (CE) US, the American College of Radiology developed the CEUS Liver Imaging Reporting and Data System (LI-RADS). However, the system still exhibits limitations in distinguishing between HCC and non-HCC lesions. PurposeTo investigate the viability of employing machine learning methods based on quantitative parameters of contrast-enhanced ultrasound for distinguishing HCC within LR-M nodules. Materials and methodsThis retrospective analysis was conducted on pre-treatment CEUS data from liver nodule patients across multiple centers between January 2013 and June 2022. Quantitative analysis was performed using CEUS images, and the machine learning diagnostic models based on quantitative parameters were utilized for the classification diagnosis of LR-M nodules. The performance of the model was assessed using the area under the receiver operating characteristic curve (AUC) and compared with the performance of four radiologists. ResultsThe training and internal testing datasets comprised 168 patients (median age, 53 years [IQR, 18 years]), while the external testing datasets from two other centers included 110 patients (median age, 54 years [IQR, 16 years]). In the internal independent test set, the top-performing Random Forest model achieved an AUC of 0.796 (95%CI: 0.729–0.853) for diagnosing HCC. This model exhibited a sensitivity of 0.752 (95%CI: 0.750–0.755) and a specificity of 0.761 (95%CI: 0.758–0.764), outperforming junior radiologists who achieved an AUC of 0.619 (95%CI: 0.543–0.691, p < .01) with sensitivity and specificity of 0.716 (95%CI: 0.713–0.718) and 0.522 (95%CI: 0.519–0.526), respectively. ConclusionSignificant differences in contrast-enhanced ultrasound quantitative parameters are observed between HCC and non-HCC lesions. Machine learning models leveraging these parameters effectively distinguish HCC categorized as LR-M, offering a valuable adjunct for the accurate classification of liver nodules within the CEUS LI-RADS framework.

An ensemble machine learning model assists in the diagnosis of gastric ectopic pancreas and gastric stromal tumors

ObjectiveTo develop an ensemble machine learning (eML) model using multiphase computed tomography (MPCT) for distinguishing between gastric ectopic pancreas (GEP) and gastric stromal tumors (GIST) in lesions < 3 cm.MethodsIn this study, we retrospectively collected MPCT images from 138 patients between April 2017 and June 2023 across two centers. Cohort 1 comprised 94 patients divided into a training cohort and an internal validation cohort, while the 44 patients from Cohort 2 constituted the external validation cohort. Deep learning (DL) models were constructed based on the lesion region, and radiomics features were extracted to develop radiomics models, which were later integrated into the fusion model. Model performance was assessed through the analysis of the area under the receiver operating characteristic curve (AUROC). The diagnostic efficacy of the optimal model was compared with that of a radiologist. Additionally, the radiologist with the assistance of the eML model provides a secondary diagnosis, to assess the potential clinical value of the model.ResultsAfter evaluation using an external validation cohort, the radiomics model demonstrated the highest performance in the venous phase, achieving AUROC of 0.87. The DL model showed optimal performance in the non-contrast phase, with AUROC of 0.81. The eML achieved the best performance across all models, with AUROC of 0.90. The use of eML-assisted analysis resulted in a significant improvement in the junior radiologist’s accuracy, rising from 0.77 to 0.93 (p < 0.05). However, the senior radiologist’s accuracy, while improving from 0.86 to 0.95, did not exhibit a statistically significant difference.ConclusioneML model based on MPCT can effectively distinguish between GEPs and GISTs < 3 cm.Critical relevance statementThe multiphase CT-based fusion model, incorporating radiomics and DL technology, proves effective in distinguishing between GEP and gastric stromal tumors, serving as a valuable tool to enhance diagnoses and offering references for clinical decision-making.Key PointsNo studies yet differentiated these tumors via radiomics or DL.Radiomics and DL methodologies unveil potentially distinct phenotypes within lesions.Quantitative analysis on CT for GIST and ectopic pancreas.Ensemble learning aids accurate diagnoses, assisting treatment decisions.Graphical

Deep learning-assisted distinguishing breast phyllodes tumours from fibroadenomas based on ultrasound images: a diagnostic study.

To evaluate the performance of ultrasound-based deep learning (DL) models in distinguishing breast phyllodes tumours (PTs) from fibroadenomas (FAs) and their clinical utility in assisting radiologists with varying diagnostic experiences. We retrospectively collected 1180 ultrasound images from 539 patients (247 PTs and 292 FAs). Five DL network models with different structures were trained and validated using nodule regions annotated by radiologists on breast ultrasound images. DL models were trained using the methods of transfer learning and 3-fold cross-validation. The model demonstrated the best evaluation index in the 3-fold cross-validation was selected for comparison with radiologists' diagnostic decisions. Two-round reader studies were conducted to investigate the value of DL model in assisting 6 radiologists with different levels of experience. Upon testing, Xception model demonstrated the best diagnostic performance (area under the receiver-operating characteristic curve: 0.87; 95% CI, 0.81-0.92), outperforming all radiologists (all P < .05). Additionally, the DL model enhanced the diagnostic performance of radiologists. Accuracy demonstrated improvements of 4%, 4%, and 3% for senior, intermediate, and junior radiologists, respectively. The DL models showed superior predictive abilities compared to experienced radiologists in distinguishing breast PTs from FAs. Utilizing the model led to improved efficiency and diagnostic performance for radiologists with different levels of experience (6-25 years of work). We developed and validated a DL model based on the largest available dataset to assist in diagnosing PTs. This model has the potential to allow radiologists to discriminate 2 types of breast tumours which are challenging to identify with precision and accuracy, and subsequently to make more informed decisions about surgical plans.

Prediction of axillary lymph node metastasis using a magnetic resonance imaging radiomics model of invasive breast cancer primary tumor

BackgroundThis study investigated the clinical value of breast magnetic resonance imaging (MRI) radiomics for predicting axillary lymph node metastasis (ALNM) and to compare the discriminative abilities of different combinations of MRI sequences.MethodsThis study included 141 patients diagnosed with invasive breast cancer from two centers (center 1: n = 101, center 2: n = 40). Patients from center 1 were randomly divided into training set and test set 1. Patients from center 2 were assigned to the test set 2. All participants underwent preoperative MRI, and four distinct MRI sequences were obtained. The volume of interest (VOI) of the breast tumor was delineated on the dynamic contrast-enhanced (DCE) postcontrast phase 2 sequence, and the VOIs of other sequences were adjusted when required. Subsequently, radiomics features were extracted from the VOIs using an open-source package. Both single- and multisequence radiomics models were constructed using the logistic regression method in the training set. The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and precision of the radiomics model for the test set 1 and test set 2 were calculated. Finally, the diagnostic performance of each model was compared with the diagnostic level of junior and senior radiologists.ResultsThe single-sequence ALNM classifier derived from DCE postcontrast phase 1 had the best performance for both test set 1 (AUC = 0.891) and test set 2 (AUC = 0.619). The best-performing multisequence ALNM classifiers for both test set 1 (AUC = 0.910) and test set 2 (AUC = 0.717) were generated from DCE postcontrast phase 1, T2-weighted imaging, and diffusion-weighted imaging single-sequence ALNM classifiers. Both had a higher diagnostic level than the junior and senior radiologists.ConclusionsThe combination of DCE postcontrast phase 1, T2-weighted imaging, and diffusion-weighted imaging radiomics features had the best performance in predicting ALNM from breast cancer. Our study presents a well-performing and noninvasive tool for ALNM prediction in patients with breast cancer.

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.

Learning co-plane attention across MRI sequences for diagnosing twelve types of knee abnormalities

Multi-sequence magnetic resonance imaging is crucial in accurately identifying knee abnormalities but requires substantial expertise from radiologists to interpret. Here, we introduce a deep learning model incorporating co-plane attention across image sequences to classify knee abnormalities. To assess the effectiveness of our model, we collected the largest multi-sequence knee magnetic resonance imaging dataset involving the most comprehensive range of abnormalities, comprising 1748 subjects and 12 types of abnormalities. Our model achieved an overall area under the receiver operating characteristic curve score of 0.812. It achieved an average accuracy of 0.78, outperforming junior radiologists (accuracy 0.65) and remains competitive with senior radiologists (accuracy 0.80). Notably, with the assistance of model output, the diagnosis accuracy of all radiologists was improved significantly (p < 0.001), elevating from 0.73 to 0.79 on average. The interpretability analysis demonstrated that the model decision-making process is consistent with the clinical knowledge, enhancing its credibility and reliability in clinical practice.

Development and validation of deep learning models for bowel obstruction on plain abdominal radiograph.

Artificial intelligence (AI) could help medical practitioners in analyzing radiological images to determine the presence and site of bowel obstruction. This retrospective diagnostic study proposed a series of deep learning (DL) models for diagnosing bowel obstruction on abdominal radiograph. A total of 2082 upright plain abdominal radiographs were retrospectively collected from four hospitals. The images were labeled as normal, small bowel obstruction and large bowel obstruction by three senior radiologists based on comprehensive examinations and interventions within 48 hours after admission. Gradient-weighted class activation mapping was used to visualize the inferential explanation. In the validation set, the Xception-backboned model achieved the highest accuracy (0.863), surpassing the VGG16 (0.847) and ResNet models (0.836). In the test set, the Xception model (accuracy: 0.807) outperformed other models and a junior radiologist (0.780) but not a senior radiologist (0.840). In the AI-aided diagnostic framework, the junior and senior radiologists made their judgements while aware of the Xception model predictions. Their accuracy significantly improved to 0.887 and 0.913, respectively. We developed and validated DL-based computer vision models for diagnosing bowel obstruction on plain abdominal radiograph. DL-based computer-aided diagnostic systems could reduce medical practitioners' workloads and improve diagnostic accuracy.

Standardized IETA criteria enhance accuracy of junior and intermediate ultrasound radiologists in diagnosing malignant endometrial and intrauterine lesions.

To transform the standardized descriptions of the ultrasound characteristics of endometrial and intrauterine lesions devised by the International Endometrial Tumor Analysis (IETA) group into a practical scoring method and to investigate whether application of this method enhances the diagnostic accuracy of ultrasound radiologists with different levels of experience in detecting malignancy compared with subjective assessment. This was a retrospective study of 855 patients with endometrial and/or intrauterine lesions, who were divided into a training (n = 600) and a validation (n = 255) set. Ultrasound radiologists with varying levels of experience (expert, intermediate and junior) evaluated all lesions by subjective assessment and according to IETA rules. Using IETA rules, the experts identified signs of malignancy in the training set, assigned scores for each indicator and validated the scoring method in the validation set. The intermediate-level and junior ultrasound radiologists reassessed the malignancy of the lesions using the IETA scoring method and compared their classifications with those made previously by subjective assessment. Postsurgical pathological evaluation was used as the reference standard. Using subjective assessment, the experts demonstrated the highest level of diagnostic accuracy, with a sensitivity of 85.0%, specificity of 94.3% and an area under the receiver-operating-characteristics curve (AUC) of 0.897. Applying the IETA scoring method (comprising eight ultrasound characteristics that contributed to the total score) with a threshold of > 25 points for the diagnosis of malignancy achieved a sensitivity of 84.7%, specificity of 94.7% and AUC of 0.9533 in the training set, with similar performance in the validation set, when performed by experts. Using the IETA scoring method, both junior and intermediate ultrasound radiologists showed improvement in sensitivity (from 55.5% to 74.8% and from 70.2% to 77.1%, respectively), specificity (from 88.4% to 91.5% and from 87.4% to 92.2%, respectively) and AUC (from 0.704 to 0.827 and from 0.793 to 0.841, respectively) for diagnosing malignant lesions. The IETA scoring method exhibits high diagnostic efficacy for malignant endometrial and intrauterine lesions. This method compensates for the lack of experience among junior and intermediate-level ultrasound radiologists, enhancing their diagnostic skill to a level nearing that of experienced senior ultrasound radiologists. Further research is essential to validate the practicality of implementing this method and to confirm its clinical value. © 2024 International Society of Ultrasound in Obstetrics and Gynecology.

A few-shot learning framework for the diagnosis of osteopenia and osteoporosis using knee X-ray images.

We developed a few-shot learning (FSL) framework for the diagnosis of osteopenia and osteoporosis in knee X-ray images. Computer vision models containing deep convolutional neural networks were fine-tuned to enable generalization from natural images (ImageNet) to chest X-ray images (normal vs. pneumonia, base images). Then, a series of automated machine learning classifiers based on the Euclidean distances of base images were developed to make predictions for novel images (normal vs. osteopenia vs. osteoporosis). The performance of the FSL framework was compared with that of junior and senior radiologists. In addition, the gradient-weighted class activation mapping algorithm was used for visual interpretation. In Cohort #1, the mean accuracy (0.728) and sensitivity (0.774) of the FSL models were higher than those of the radiologists (0.512 and 0.448). A diagnostic pipeline of FSL model (first)-radiologists (second) achieved better performance (0.653 accuracy, 0.582 sensitivity, and 0.816 specificity) than radiologists alone. In Cohort #2, the diagnostic pipeline also showed improved performance. The FSL framework yielded practical performance with respect to the diagnosis of osteopenia and osteoporosis in comparison with radiologists. This retrospective study supports the use of promising FSL methods in computer-aided diagnosis tasks involving limited samples.