Junior Radiologists Research Articles

Improved breast lesion detection in mammogram images using a deep neural network.

This study aimed to investigate the effect of using a deep neural network (DNN) in breast cancer (BC) detection. In this retrospective study, a DNN-based model was constructed from a total of 880 mammograms that 220 patients underwent between April and June 2020. The mammograms were reviewed by two senior and two junior radiologists with and without the aid of the DNN model. The performance of the network was assessed by comparing the area under the curve (AUC) and receiver operating characteristic curves for the detection of four features of malignancy (masses, calcifications, asymmetries, and architectural distortions), with and without the aid of the DNN model and by the senior and junior radiologists. Additionally, the effect of utilizing the DNN on diagnosis time for both the senior and junior radiologists was evaluated. The AUCs of the model for the detection of mass and calcification were 0.877 and 0.937, respectively. In the senior radiologist group, the AUC values for evaluation of mass, calcification, and asymmetric compaction were significantly higher with the DNN model than those obtained without the model. Similar effects were observed in the junior radiologist group, but the increase in the AUC values was even more dramatic. The median mammogram assessment time of the junior and senior radiologists was 572 (357-951) s, and 273.5 (129-469) s, respectively, with the DNN model, and the corresponding assessment time without the model, was 739 (445-1003) s and 321 (195-491) s, respectively. The DNN model exhibited high accuracy in detecting the four named features of BC and effectively shortened the review time by both senior and junior radiologists.

Uterine mesenchymal tumors: development and preliminary results of a magnetic resonance imaging (MRI) diagnostic algorithm.

The aim of our study is to propose a diagnostic algorithm to guide MRI findings interpretation and malignancy risk stratification of uterine mesenchymal masses with a multiparametric step-by-step approach. A non-interventional retrospective multicenter study was performed: Preoperative MRI of 54 uterine masses was retrospectively evaluated. Firstly, the performance of MRI with monoparametric and multiparametric approach was assessed. Reference standard for final diagnosis was surgical pathologic result (n = 53 patients) or at least 1-year MR imaging follow-up (n = 1 patient). Subsequently, a diagnostic algorithm was developed for MR interpretation, resulting in a Likert score from 1 to 5 predicting risk of malignancy of the uterine lesion. The accuracy and reproducibility of the MRI scoring system were then tested: 26 preoperative pelvic MRI were double-blind evaluated by a senior (SR) and junior radiologist (JR). Diagnostic performances and the agreement between the two readers with and without the application of the proposed algorithm were compared, using histological results as standard reference. Multiparametric approach showed the best diagnostic performance in terms of accuracy (94.44%,) and specificity (97.56%). DWI was confirmed as the most sensible parameter with a relative high specificity: low ADC values (mean 0.66) significantly correlated to uterine sarcomas diagnosis (p < 0.01). Proposed algorithm allowed to improve both JR and SR performance (algorithm-aided accuracy 88.46% and 96%, respectively) and determined a significant increase in inter-observer agreement, helping even the less-experienced radiologist in this difficult differential diagnosis. Uterine leiomyomas and sarcomas often show an overlap of clinical and imaging features. The application of a diagnostic algorithm can help radiologists to standardize their approach to a complex myometrial mass and to easily identify suspicious MRI features favoring malignancy.

Automatic Image Segmentation and Grading Diagnosis of Sacroiliitis Associated with AS Using a Deep Convolutional Neural Network on CT Images.

Ankylosing spondylitis (AS) is a chronic inflammatory disease that causes inflammatory low back pain and may even limit activity. The grading diagnosis of sacroiliitis on imaging plays a central role in diagnosing AS. However, the grading diagnosis of sacroiliitis on computed tomography (CT) images is viewer-dependent and may vary between radiologists and medical institutions. In this study, we aimed to develop a fully automatic method to segment sacroiliac joint (SIJ) and further grading diagnose sacroiliitis associated with AS on CT. We studied 435 CT examinations from patients with AS and control at two hospitals. No-new-UNet (nnU-Net) was used to segment the SIJ, and a 3D convolutional neural network (CNN) was used to grade sacroiliitis with a three-class method, using the grading results of three veteran musculoskeletal radiologists as the ground truth. We defined grades 0-I as class 0, grade II as class 1, and grades III-IV as class 2 according to modified New York criteria. nnU-Net segmentation of SIJ achieved Dice, Jaccard, and relative volume difference (RVD) coefficients of 0.915, 0.851, and 0.040 with the validation set, respectively, and 0.889, 0.812, and 0.098 with the test set, respectively. The areas under the curves (AUCs) of classes 0, 1, and 2 using the 3D CNN were 0.91, 0.80, and 0.96 with the validation set, respectively, and 0.94, 0.82, and 0.93 with the test set, respectively. 3D CNN was superior to the junior and senior radiologists in the grading of class 1 for the validation set and inferior to expert for the test set (P < 0.05). The fully automatic method constructed in this study based on a convolutional neural network could be used for SIJ segmentation and then accurately grading and diagnosis of sacroiliitis associated with AS on CT images, especially for class 0 and class 2. The method for class 1 was less effective but still more accurate than that of the senior radiologist.

Generalizability and Diagnostic Performance of AI Models for Thyroid US.

Background Artificial intelligence (AI) models have improved US assessment of thyroid nodules; however, the lack of generalizability limits the application of these models. Purpose To develop AI models for segmentation and classification of thyroid nodules in US using diverse data sets from nationwide hospitals and multiple vendors, and to measure the impact of the AI models on diagnostic performance. Materials and Methods This retrospective study included consecutive patients with pathologically confirmed thyroid nodules who underwent US using equipment from 12 vendors at 208 hospitals across China from November 2017 to January 2019. The detection, segmentation, and classification models were developed based on the subset or complete set of images. Model performance was evaluated by precision and recall, Dice coefficient, and area under the receiver operating characteristic curve (AUC) analyses. Three scenarios (diagnosis without AI assistance, with freestyle AI assistance, and with rule-based AI assistance) were compared with three senior and three junior radiologists to optimize incorporation of AI into clinical practice. Results A total of 10 023 patients (median age, 46 years [IQR 37-55 years]; 7669 female) were included. The detection, segmentation, and classification models had an average precision, Dice coefficient, and AUC of 0.98 (95% CI: 0.96, 0.99), 0.86 (95% CI: 0.86, 0.87), and 0.90 (95% CI: 0.88, 0.92), respectively. The segmentation model trained on the nationwide data and classification model trained on the mixed vendor data exhibited the best performance, with a Dice coefficient of 0.91 (95% CI: 0.90, 0.91) and AUC of 0.98 (95% CI: 0.97, 1.00), respectively. The AI model outperformed all senior and junior radiologists (P < .05 for all comparisons), and the diagnostic accuracies of all radiologists were improved (P < .05 for all comparisons) with rule-based AI assistance. Conclusion Thyroid US AI models developed from diverse data sets had high diagnostic performance among the Chinese population. Rule-based AI assistance improved the performance of radiologists in thyroid cancer diagnosis. © RSNA, 2023 Supplemental material is available for this article.

CT Radiomics-Based Nomogram for Predicting the Lateral Neck Lymph Node Metastasis in Papillary Thyroid Carcinoma: A Prospective Multicenter Study

This study is based on multicenter cohorts and aims to utilize computed tomography (CT) images to construct a radiomics nomogram for predicting the lateral neck lymph node (LNLN) metastasis in the papillary thyroid carcinoma (PTC) and further explore the biological basis under its prediction. In the multicenter study, 1213 lymph nodes from 409 patients with PTC who underwent CT examinations and received open surgery and lateral neck dissection were included. A prospective test cohort was used in validating the model. Radiomics features were extracted from the CT images of each patient's LNLNs. Selectkbest, maximum relevance and minimum redundancy and the least absolute shrinkage and selection operator (LASSO) algorithm were used in reducing the dimensionality of radiomics features in the training cohort. Then, a radiomics signature (Rad-score) was calculated as the sum of each feature multiplied by the nonzero coefficient from LASSO. A nomogram was generated using the clinical risk factors of the patients and Rad-score. The nomograms' performance was analyzed in terms of accuracy, sensitivity, specificity, confusion matrix, receiver operating characteristic curves, and areas under the receiver operating characteristic curve (AUCs). The clinical usefulness of the nomogram was evaluated by decision curve analysis. Moreover, three radiologists with different working experiences and nomogram were compared to one another. Whole transcriptomics sequencing was performed in 14 tumor samples; the correlation of biological functions and high and low LNLN samples predicted by the nomogram was further investigated. A total of 29 radiomics features were used in constructing the Rad-score. Rad-score and clinical risk factors (age, tumor diameter, location and number of suspected tumors) compose the nomogram. The nomogram exhibited good discrimination performance of the nomogram for predicting LNLN metastasis in the training cohort (AUC, 0.866), internal test cohort (0.845), external test cohort (0.725), and prospective test cohort (0.808) and showed diagnostic capability comparable to senior radiologists, significantly outperforming junior radiologists (p<0.05). Functional enrichment analysis suggested that the nomogram can reflect the ribosome-related structures of cytoplasmic translation in patients with PTC. Our radiomics nomogram provides a noninvasive method that incorporates radiomics features and clinical risk factors for predicting LNLN metastasis in patients with PTC.

Integration of Artificial Intelligence Decision Aids to Reduce Workload and Enhance Efficiency in Thyroid Nodule Management

To optimize the integration of artificial intelligence (AI) decision aids and reduce workload in thyroid nodule management, it is critical to incorporate personalized AI into the decision-making processes of radiologists with varying levels of expertise. To develop an optimized integration of AI decision aids for reducing radiologists' workload while maintaining diagnostic performance compared with traditional AI-assisted strategy. In this diagnostic study, a retrospective set of 1754 ultrasonographic images of 1048 patients with 1754 thyroid nodules from July 1, 2018, to July 31, 2019, was used to build an optimized strategy based on how 16 junior and senior radiologists incorporated AI-assisted diagnosis results with different image features. In the prospective set of this diagnostic study, 300 ultrasonographic images of 268 patients with 300 thyroid nodules from May 1 to December 31, 2021, were used to compare the optimized strategy with the traditional all-AI strategy in terms of diagnostic performance and workload reduction. Data analyses were completed in September 2022. The retrospective set of images was used to develop an optimized integration of AI decision aids for junior and senior radiologists based on the selection of AI-assisted significant or nonsignificant features. In the prospective set of images, the diagnostic performance, time-based cost, and assisted diagnosis were compared between the optimized strategy and the traditional all-AI strategy. The retrospective set included 1754 ultrasonographic images from 1048 patients (mean [SD] age, 42.1 [13.2] years; 749 women [71.5%]) with 1754 thyroid nodules (mean [SD] size, 16.4 [10.6] mm); 748 nodules (42.6%) were benign, and 1006 (57.4%) were malignant. The prospective set included 300 ultrasonographic images from 268 patients (mean [SD] age, 41.7 [14.1] years; 194 women [72.4%]) with 300 thyroid nodules (mean [SD] size, 17.2 [6.8] mm); 125 nodules (41.7%) were benign, and 175 (58.3%) were malignant. For junior radiologists, the ultrasonographic features that were not improved by AI assistance included cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, and nodules smaller than 5 mm, whereas for senior radiologists the features that were not improved by AI assistance were cystic or almost completely cystic nodules, anechoic nodules, spongiform nodules, very hypoechoic nodules, nodules taller than wide, lobulated or irregular nodules, and extrathyroidal extension. Compared with the traditional all-AI strategy, the optimized strategy was associated with increased mean task completion times for junior radiologists (reader 11, from 15.2 seconds [95% CI, 13.2-17.2 seconds] to 19.4 seconds [95% CI, 15.6-23.3 seconds]; reader 12, from 12.7 seconds [95% CI, 11.4-13.9 seconds] to 15.6 seconds [95% CI, 13.6-17.7 seconds]), but shorter times for senior radiologists (reader 14, from 19.4 seconds [95% CI, 18.1-20.7 seconds] to 16.8 seconds [95% CI, 15.3-18.3 seconds]; reader 16, from 12.5 seconds [95% CI, 12.1-12.9 seconds] to 10.0 seconds [95% CI, 9.5-10.5 seconds]). There was no significant difference in sensitivity (range, 91%-100%) or specificity (range, 94%-98%) between the 2 strategies for readers 11 to 16. This diagnostic study suggests that an optimized AI strategy in thyroid nodule management may reduce diagnostic time-based costs without sacrificing diagnostic accuracy for senior radiologists, while the traditional all-AI strategy may still be more beneficial for junior radiologists.

Collimation border with U-Net segmentation on chest radiographs compared to radiologists

IntroductionChest Radiography (CXR) is a common radiographic procedure. Radiation exposure to patients should be kept as low as reasonably achievable (ALARA), and monitored continuously as part of quality assurance (QA) programs. One of the most effective dose reduction tools is proper collimation practice.The purpose of this study is to determine whether a U-Net convolutional neural networks (U-CNN) can be trained to automatically segment the lungs and calculate an optimized collimation border on a limited CXR dataset. Methods662 CXRs with manual lung segmentations were obtained from an open-source dataset. These were used to train and validate three different U-CNNs for automatic lung segmentation and optimal collimation. The U-CNN dimensions were 128 × 128, 256 × 256, and 512 × 512 pixels and validated with five-fold cross validation. The U-CNN with the highest area under the curve (AUC) was tested externally, using a dataset of 50 CXRs. Dice scores (DS) were used to compare U-CNN segmentations with manual segmentations by three radiographers and two junior radiologists. ResultsDS for the three U-CNN dimensions with segmentation of the lungs ranged from 0.93 to 0.96, respectively. DS of the collimation border for each U-CNN was 0.95 compared to the ground truth labels. DS for lung segmentation and collimation border between the junior radiologists was 0.97 and 0.97. One radiographer differed significantly from the U-CNN (p = 0.016). ConclusionWe demonstrated that a U-CNN could reliably segment the lungs and suggest a collimation border with great accuracy compared to junior radiologists. This algorithm has the potential to automate collimation auditing of CXRs. Implications for practiceCreating an automatic segmentation model of the lungs can produce a collimation border, which can be used in CXR QA programs.

Machine learning combined with radiomics and deep learning features extracted from CT images: a novel AI model to distinguish benign from malignant ovarian tumors

BackgroundTo develop an artificial intelligence (AI) model with radiomics and deep learning (DL) features extracted from CT images to distinguish benign from malignant ovarian tumors.MethodsWe enrolled 149 patients with pathologically confirmed ovarian tumors. A total of 185 tumors were included and divided into training and testing sets in a 7:3 ratio. All tumors were manually segmented from preoperative contrast-enhanced CT images. CT image features were extracted using radiomics and DL. Five models with different combinations of feature sets were built. Benign and malignant tumors were classified using machine learning (ML) classifiers. The model performance was compared with five radiologists on the testing set.Results Among the five models, the best performing model is the ensemble model with a combination of radiomics, DL, and clinical feature sets. The model achieved an accuracy of 82%, specificity of 89% and sensitivity of 68%. Compared with junior radiologists averaged results, the model had a higher accuracy (82% vs 66%) and specificity (89% vs 65%) with comparable sensitivity (68% vs 67%). With the assistance of the model, the junior radiologists achieved a higher average accuracy (81% vs 66%), specificity (80% vs 65%), and sensitivity (82% vs 67%), approaching to the performance of senior radiologists.Conclusions We developed a CT-based AI model that can differentiate benign and malignant ovarian tumors with high accuracy and specificity. This model significantly improved the performance of less-experienced radiologists in ovarian tumor assessment, and may potentially guide gynecologists to provide better therapeutic strategies for these patients.

Automated detection and classification of acute vertebral body fractures using a convolutional neural network on computed tomography.

Acute vertebral fracture is usually caused by low-energy injury with osteoporosis and high-energy trauma. The AOSpine thoracolumbar spine injury classification system (AO classification) plays an important role in the diagnosis and treatment of the disease. The diagnosis and description of vertebral fractures according to the classification scheme requires a great deal of time and energy for radiologists. To design and validate a multistage deep learning system (multistage AO system) for the automatic detection, localization and classification of acute thoracolumbar vertebral body fractures according to AO classification on computed tomography. The CT images of 1,217 patients who came to our hospital from January 2015 to December 2019 were collected retrospectively. The fractures were marked and classified by 2 junior radiology residents according to the type A standard in the AO classification. Marked fracture sites included the upper endplate, lower endplate and posterior wall. When there were inconsistent opinions on classification labels, the final result was determined by a director radiologist. We integrated different networks into different stages of the overall framework. U-net and a graph convolutional neural network (U-GCN) are used to realize the location and classification of the thoracolumbar spine. Next, a classification network is used to detect whether the thoracolumbar spine has a fracture. In the third stage, we detect fractures in different parts of the thoracolumbar spine by using a multibranch output network and finally obtain the AO types. The mean age of the patients was 61.87 years with a standard deviation of 17.04 years, consisting of 760 female patients and 457 male patients. On vertebrae level, sensitivity for fracture detection was 95.23% in test dataset, with an accuracy of 97.93% and a specificity of 98.35%. For the classification of vertebral body fractures, the balanced accuracy was 79.56%, with an AUC of 0.904 for type A1, 0.945 for type A2, 0.878 for type A3 and 0.942 for type A4. The multistage AO system can automatically detect and classify acute vertebral body fractures in the thoracolumbar spine on CT images according to AO classification with high accuracy.

Ultrasound image-based deep learning to assist in diagnosing gross extrathyroidal extension thyroid cancer: a retrospective multicenter study

The presence of gross extrathyroidal extension (ETE) in thyroid cancer will affect the prognosis of patients, but imaging examination cannot provide a reliable diagnosis for it. This study was conducted to develop a deep learning (DL) model for localization and evaluation of thyroid cancer nodules in ultrasound images before surgery for the presence of gross ETE. From January 2016 to December 2021 grayscale ultrasound images of 806 thyroid cancer nodules (4451 images) from 4 medical centers were retrospectively analyzed, including 517 no gross ETE nodules and 289 gross ETE nodules. 283 no gross ETE nodules and 158 gross ETE nodules were randomly selected from the internal dataset to form a training set and validation set (2914 images), and a multitask DL model was constructed for diagnosing gross ETE. In addition, the clinical model and the clinical and DL combined model were constructed. In the internal test set [974 images (139 no gross ETE nodules and 83 gross ETE nodules)] and the external test set [563 images (95 no gross ETE nodules and 48 gross ETE nodules)], the diagnostic performance of DL model was verified based on the pathological results. And then, compared the results with the diagnosis by 2 senior and 2 junior radiologists. In the internal test set, DL model demonstrated the highest AUC (0.91; 95% CI: 0.87, 0.96), which was significantly higher than that of two senior radiologists [(AUC, 0.78; 95% CI: 0.71, 0.85; P<0.001) and (AUC, 0.76; 95% CI: 0.70, 0.83; P<0.001)] and two juniors radiologists [(AUC, 0.65; 95% CI: 0.58, 0.73; P<0.001) and (AUC, 0.69; 95% CI: 0.62, 0.77; P<0.001)]. DL model was significantly higher than clinical model [(AUC, 0.84; 95% CI: 0.79, 0.89; P=0.019)], but there was no significant difference between DL model and clinical and DL combined model [(AUC, 0.94; 95% CI: 0.91, 0.97; P=0.143)]. In the external test set, DL model also demonstrated the highest AUC (0.88, 95% CI: 0.81, 0.94), which was significantly higher than that of one of senior radiologists [(AUC, 0.75; 95% CI: 0.66, 0.84; P=0.008) and (AUC, 0.81; 95% CI: 0.72, 0.89; P=0.152)] and two junior radiologists [(AUC, 0.72; 95% CI: 0.62, 0.81; P=0.002) and (AUC, 0.67; 95 CI: 0.57, 0.77; P<0.001]. There was no significant difference between DL model and clinical model [(AUC, 0.85; 95% CI: 0.79, 0.91; P=0.516)] and clinical+DL model [(AUC, 0.92; 95% CI: 0.87, 0.96; P=0.093)]. Using DL model, the diagnostic ability of two junior radiologists was significantly improved. The DL model based on ultrasound imaging is a simple and helpful tool for preoperative diagnosis of gross ETE thyroid cancer, and its diagnostic performance is equivalent to or even better than that of senior radiologists. Jiangxi Provincial Natural Science Foundation (20224BAB216079), the Key Research and Development Program of Jiangxi Province (20181BBG70031), and the Interdisciplinary Innovation Fund of Natural Science, Nanchang University (9167-28220007-YB2110).

Predicting prostate cancer in men with PSA levels of 4–10 ng/mL: MRI-based radiomics can help junior radiologists improve the diagnostic performance

To develop MRI-based radiomics model for predicting prostate cancer (PCa) in men with prostate-specific antigen (PSA) levels of 4–10 ng/mL, to compare the performance of radiomics model and PI-RADS v2.1, and to further verify the predictive ability of radiomics model for lesions with different PI-RADS v2.1 score. 171 patients with PSA levels of 4–10 ng/mL were divided into training (n = 119) and testing (n = 52) groups. PI-RADS v2.1 score was assessed by two radiologists. All volumes of interest were segmented on T2-weighted imaging, diffusion weighted imaging, and apparent diffusion coefficient sequences, from which quantitative radiomics features were extracted. Multivariate logistic regression analysis was performed to establish radiomics model for predicting PCa. The diagnostic performance was assessed using receiver operating characteristic curve analysis. The radiomics model exhibited the best performance in predicting PCa, which was better than the performance of PI-RADS v2.1 scoring by the junior radiologist in the training group [area under the curve (AUC): 0.932 vs 0.803], testing group (AUC: 0.922 vs 0.797), and the entire cohort (AUC: 0.927 vs 0.801) (P < 0.05). The radiomics model performed well for lesions with PI-RADS v2.1 score of 3 (AUC = 0.854, sensitivity = 84.62%, specificity = 84.34%) and PI-RADS v2.1 score of 4–5 (AUC = 0.967, sensitivity = 98.11%, specificity = 86.36%) assigned by junior radiologist. The radiomics model quantitatively outperformed PI-RADS v2.1 for noninvasive prediction of PCa in men with PSA levels of 4–10 ng/mL. The model can help improve the diagnostic performance of junior radiologists and facilitate better decision-making by urologists for management of lesions with different PI-RADS v2.1 score.

Automatic deep learning-based assessment of spinopelvic coronal and sagittal alignment

PurposeThe purpose of this study was to evaluate an artificial intelligence (AI) solution for estimating coronal and sagittal spinopelvic alignment on conventional uniplanar two-dimensional whole-spine radiograph. Material and methodsThis retrospective observational study included 100 patients (35 men, 65 women) with a median age of 14 years (IQR: 13, 15.25; age range: 3–64 years) who underwent conventional uniplanar two-dimensional whole-spine radiograph in standing position between January and July 2022. Ten most commonly used spinopelvic coronal and sagittal parameters were retrospectively measured without AI by a junior radiologist and approved or adjusted by a senior musculoskeletal radiologist to reach final measurements. Final measurements were used as the ground truth to assess AI performance for each parameter. AI performances were estimated using mean absolute errors (MAE), intraclass correlation coefficient (ICCs), and accuracy for selected clinically relevant thresholds. Readers visually classified AI outputs to assess reliability at a patient-level. ResultsAI solution showed excellent consistency without bias in coronal (ICCs ≥ 0.95; MAE ≤ 2.9° or 1.97 mm) and sagittal (ICCs ≥ 0.85; MAE ≤ 4.4° or 2.7 mm) spinopelvic evaluation, except for kyphosis (ICC = 0.58; MAE = 8.7°). AI accuracy to classify low Cobb angle, severe scoliosis or frontal pelvic asymmetry was 91% (95% CI: 85–96), 99% (95% CI: 97–100) and 94% (95% CI: 89–98), respectively. Overall, AI provided reliable measurements in 72/100 patients (72%). ConclusionThe AI solution used in this study for combined coronal and sagittal spinopelvic balance assessment provides results consistent with those of a senior musculoskeletal radiologist, and shows potential benefit for reducing workload in future routine implementation.

USC-ENet: a high-efficiency model for the diagnosis of liver tumors combining B-mode ultrasound and clinical data.

Ultrasound image acquisition has the advantages of being low cost, rapid, and non-invasive, and it does not produce radiation. Currently, ultrasound is widely used in the diagnosis of liver tumors. However, owing to the complex presentation and diverse features of benign and malignant liver tumors, accurate diagnosis of liver tumors using ultrasound is difficult even for experienced radiologists. In recent years, artificial intelligence-assisted diagnosis has proven to provide effective support to radiologists. However, there is room for further improvement in the existing ultrasound artificial intelligence diagnostic model of liver tumor. First, the image diagnostic model may not fully consider relevant clinical data in the decision-making process. Second, owing to the difficulty in collecting biopsy pathology and physician-labeled ultrasound data of liver tumors, training datasets are usually small, and commonly used large neural networks tend to overfit on small datasets, which seriously affects the generalization of the model. In this study, we propose a deep learning-assisted diagnosis model called USC-ENet, which integrates B-mode ultrasound features of liver tumors and clinical data of patients, and we design a small neural network specifically for small-scale medical images combined with an attention mechanism. Real data from 542 patients with liver tumors (N = 2168 images) are used during model training and validation. Experiments show that USC-ENet can achieve a good classification effect (area under the curve = 0.956, sensitivity = 0.915, and specificity = 0.880) after small-scale data training, and it has certain interpretability, showing good potential for clinical adoption. In conclusion, our model provides not only a reliable second opinion for radiologists but also a reference for junior radiologists who lack clinical experience.

Development and clinical utility analysis of a prostate zonal segmentation model on T2-weighted imaging: a multicenter study

ObjectivesTo automatically segment prostate central gland (CG) and peripheral zone (PZ) on T2-weighted imaging using deep learning and assess the model’s clinical utility by comparing it with a radiologist annotation and analyzing relevant influencing factors, especially the prostate zonal volume.MethodsA 3D U-Net-based model was trained with 223 patients from one institution and tested using one internal testing group (n = 93) and two external testing datasets, including one public dataset (ETDpub, n = 141) and one private dataset from two centers (ETDpri, n = 59). The Dice similarity coefficients (DSCs), 95th Hausdorff distance (95HD), and average boundary distance (ABD) were calculated to evaluate the model’s performance and further compared with a junior radiologist’s performance in ETDpub. To investigate factors influencing the model performance, patients’ clinical characteristics, prostate morphology, and image parameters in ETDpri were collected and analyzed using beta regression.ResultsThe DSCs in the internal testing group, ETDpub, and ETDpri were 0.909, 0.889, and 0.869 for CG, and 0.844, 0.755, and 0.764 for PZ, respectively. The mean 95HD and ABD were less than 7.0 and 1.3 for both zones. The U-Net model outperformed the junior radiologist, having a higher DSC (0.769 vs. 0.706) and higher intraclass correlation coefficient for volume estimation in PZ (0.836 vs. 0.668). CG volume and Magnetic Resonance (MR) vendor were significant influencing factors for CG and PZ segmentation.ConclusionsThe 3D U-Net model showed good performance for CG and PZ auto-segmentation in all the testing groups and outperformed the junior radiologist for PZ segmentation. The model performance was susceptible to prostate morphology and MR scanner parameters.

Comparison of diagnostic performance of a deep learning algorithm, emergency physicians, junior radiologists and senior radiologists in the detection of appendicular fractures in children.

Advances have been made in the use of artificial intelligence (AI) in the field of diagnostic imaging, particularly in the detection of fractures on conventional radiographs. Studies looking at the detection of fractures in the pediatric population are few. The anatomical variations and evolution according to the child's age require specific studies of this population. Failure to diagnose fractures early in children may lead to serious consequences for growth. To evaluate the performance of an AI algorithm based on deep neural networks toward detecting traumatic appendicular fractures in a pediatric population. To compare sensitivity, specificity, positive predictive value and negative predictive value of different readers and the AI algorithm. This retrospective study conducted on 878 patients younger than 18years of age evaluated conventional radiographs obtained after recent non-life-threatening trauma. All radiographs of the shoulder, arm, elbow, forearm, wrist, hand, leg, knee, ankle and foot were evaluated. The diagnostic performance of a consensus of radiology experts in pediatric imaging (reference standard) was compared with those of pediatric radiologists, emergency physicians, senior residents and junior residents. The predictions made by the AI algorithm and the annotations made by the different physicians were compared. The algorithm predicted 174 fractures out of 182, corresponding to a sensitivity of 95.6%, a specificity of 91.64% and a negative predictive value of 98.76%. The AI predictions were close to that of pediatric radiologists (sensitivity 98.35%) and that of senior residents (95.05%) and were above those of emergency physicians (81.87%) and junior residents (90.1%). The algorithm identified 3 (1.6%) fractures not initially seen by pediatric radiologists. This study suggests that deep learning algorithms can be useful in improving the detection of fractures in children.

Deep learning-based algorithm improves radiologists' performance in lung cancer bone metastases detection on computed tomography.

To develop and assess a deep convolutional neural network (DCNN) model for the automatic detection of bone metastases from lung cancer on computed tomography (CT). In this retrospective study, CT scans acquired from a single institution from June 2012 to May 2022 were included. In total, 126 patients were assigned to a training cohort (n = 76), a validation cohort (n = 12), and a testing cohort (n = 38). We trained and developed a DCNN model based on positive scans with bone metastases and negative scans without bone metastases to detect and segment the bone metastases of lung cancer on CT. We evaluated the clinical efficacy of the DCNN model in an observer study with five board-certified radiologists and three junior radiologists. The receiver operator characteristic curve was used to assess the sensitivity and false positives of the detection performance; the intersection-over-union and dice coefficient were used to evaluate the segmentation performance of predicted lung cancer bone metastases. The DCNN model achieved a detection sensitivity of 0.894, with 5.24 average false positives per case, and a segmentation dice coefficient of 0.856 in the testing cohort. Through the radiologists-DCNN model collaboration, the detection accuracy of the three junior radiologists improved from 0.617 to 0.879 and the sensitivity from 0.680 to 0.902. Furthermore, the mean interpretation time per case of the junior radiologists was reduced by 228 s (p = 0.045). The proposed DCNN model for automatic lung cancer bone metastases detection can improve diagnostic efficiency and reduce the diagnosis time and workload of junior radiologists.

An Artificial Neural Network for Nasogastric Tube Position Decision Support.

To develop and validate a deep learning model for detection of nasogastric tube (NGT) malposition on chest radiographs and assess model impact as a clinical decision support tool for junior physicians to help determine whether feeding can be safely performed in patients (feed/do not feed). A neural network ensemble was pretrained on 1 132 142 retrospectively collected (June 2007-August 2019) frontal chest radiographs and further fine-tuned on 7081 chest radiographs labeled by three radiologists. Clinical relevance was assessed on an independent set of 335 images. Five junior emergency medicine physicians assessed chest radiographs and made feed/do not feed decisions without and with artificial intelligence (AI)-generated NGT malposition probabilities placed above chest radiographs. Decisions from the radiologists served as ground truths. Model performance was evaluated using receiver operating characteristic analysis. Agreement between junior physician and radiologist decision was determined using the Cohen κ coefficient. In the testing set, the ensemble achieved area under the receiver operating characteristic curve values of 0.82 (95% CI: 0.78, 0.86), 0.77 (95% CI: 0.71, 0.83), and 0.98 (95% CI: 0.96, 1.00) for satisfactory, malpositioned, and bronchial positions, respectively. In the clinical evaluation set, mean interreader agreement for feed/do not feed decisions among junior physicians was 0.65 ± 0.03 (SD) and 0.77 ± 0.13 without and with AI support, respectively. Mean agreement between junior physicians and radiologists was 0.53 ± 0.05 (unaided) and 0.65 ± 0.09 (AI-aided). A simple classifier for NGT malposition may help junior physicians determine the safety of feeding in patients with NGTs.Keywords: Neural Networks, Feature Detection, Supervised Learning, Machine Learning Supplemental material is available for this article. Published under a CC BY 4.0 license.

P337 Developing Preliminary MRI-based Classifier for Perianal Fistulizing Crohn’s Disease by Using Deep Convolutional Neural Networks

Abstract Background Early diagnosis and management of perianal fistulizing Crohn’s disease (PFCD) is critical. But there is still lack of single reliable diagnostic modality to differentiate between early PFCD and cryptoglandular anal fistula (CAF). This study aimed to evaluate the feasibility and efficacy of deep convolutional neural networks (DCNNs) for distinguishing PFCD from CAF based on pelvic magnetic resonance imaging (MRI). Methods The MRIs from 400 patients primarily diagnosed with PFCD or CAF (two hundred respectively) were retrospectively collected as datasets and used in this study. All patients had no fistula associated surgical histories. The datasets were split into training (80%), validation (10%), and test(10%). Four different DCNNs, MobileNetV2, VGG11, ResNet18 and ResNet34, were trained to classify the patients with fistula MRIs as PFCD or as CAF. Both untrained and pretrained networks were used. Pretrained networks were obtained from the Pytorch Models, an open-access repository of pretrained models for use with Pytorch. The protocol was shown in Figure 1 and was approved by the institutional review board (2022ZSLYEC-421).On the test dataset, receiver operating characteristic curves and areas under the curve (AUCs) were used to assess model performances and were compared with the performances of radiologists. Comparisons between AUCs were made by using the Delong method. Accuracy, sensitivity, and specificity were determined from the optimal threshold by the Youden index. Figure 1. Results Compared with CAF, the PFCD dataset contained more high fistulas (59.5% vs. 44.5%，P=0.003), more fistulas involved in the deep anterior space (15.5% vs. 3.0%，P&amp;lt;0.001), and less abscess formation (32.5% vs. 59.5%，P&amp;lt;0.001). Other fistula characteristics including ramification formation, supra- or infralevator extension, were not statistically different. The performances of the pretrained models were better than that of the untrained models (Figure 2). The AUC of 4 pretrained DCNN classifiers were MobileNetV2 [AUC: 0.943，95%CI（0.820~0.991）], VGG11 [AUC: 0.935, 95%CI（0.810~0.988）], ResNet 18 [AUC: 0.920, 95%CI（0.789~0.982）], ResNet 34[AUC: 0.929, 95%CI（0.801~0.986）] respectively. The performances of 4 pretrained DCNN classifiers were equivalent to that of senior radiologist, and were superior to that of junior radiologist (Figure 2, 3, 4). Figure 2. Figure 3. Figure 4. Conclusion Deep learning with DCNN in classifying PFCD or CAF on perianal MRI is feasible. Transfer learning may further improve the performance of the DCNN model. A larger sample size dataset to train the former DCCNs is conducting in our constitution and the external validation will be added in the future.

Artificial intelligence breast ultrasound and handheld ultrasound in the BI-RADS categorization of breast lesions: A pilot head to head comparison study in screening program.

Artificial intelligence breast ultrasound diagnostic system (AIBUS) has been introduced as an alternative approach for handheld ultrasound (HHUS), while their results in BI-RADS categorization has not been compared. This pilot study was based on a screening program conducted from May 2020 to October 2020 in southeast China. All the participants who received both HHUS and AIBUS were included in the study (N = 344). The ultrasound videos after AIBUS scanning were independently watched by a senior radiologist and a junior radiologist. Agreement rate and weighted Kappa value were used to compare their results in BI-RADS categorization with HHUS. The detection rate of breast nodules by HHUS was 14.83%, while the detection rates were 34.01% for AIBUS videos watched by a senior radiologist and 35.76% when watched by a junior radiologist. After AIBUS scanning, the weighted Kappa value for BI-RADS categorization between videos watched by senior radiologists and HHUS was 0.497 (p < 0.001) with an agreement rate of 78.8%, indicating its potential use in breast cancer screening. However, the Kappa value of AIBUS videos watched by junior radiologist was 0.39, when comparing to HHUS. AIBUS breast scan can obtain relatively clear images and detect more breast nodules. The results of AIBUS scanning watched by senior radiologists are moderately consistent with HHUS and might be used in screening practice, especially in primary health care with limited numbers of radiologists.

Sacroiliac joint involvement in osteochondromatosis: identifying its prevalence and characteristics from cross-sectional imaging

Apart from a few case reports, sacroiliac joint (SIJ) involvement in osteochondromatosis has not been studied. We aimed to determine the prevalence and characteristics of such involvement using cross-sectional imaging. In this retrospective study, three observers (one junior radiologist and two musculoskeletal radiologists) independently reviewed computed tomography (CT) or magnetic resonance imaging (MRI) of patients in our database who had osteochondromatosis (≥2 osteochondromas across the skeleton) for SIJ involvement. The final decision was reached by the consensus of the two musculoskeletal radiologists in a later joint session. Of the 36 patients with osteochondromatosis in our database, 22 (61%) had cross-sectional imaging covering SIJs (14 females, 8 males; age range 7-66 years; mean age 23 years; 13 MRI, 9 CT). Of these, 16 (73%) had intra-articular osteochondromas. For identifying SIJ osteochondromas on cross-sectional imaging, interobserver agreement was substantial [κ = 0.67; 95% confidence interval (CI): 0.34, 1.00] between the musculoskeletal radiologists and moderate (κ = 0.59; 95% CI: 0.23, 0.94) between the junior radiologist and the final consensus decision of the two musculoskeletal radiologists. In the cohort with cross-sectional imaging, the anatomical variations of the accessory SIJ (n = 6, 27%) and iliosacral complex (n = 2, 9%) were identified in six different patients with (n = 2) and without (n = 4) sacroiliac osteochondromas. Cross-sectional imaging shows frequent (73%) SIJ involvement in osteochondromatosis, which, although a rare disorder, nevertheless needs to be considered in the differential diagnosis of such SIJ anatomical variants as the accessory SIJ and iliosacral complex. Differentiating these variants from osteochondromas is challenging in patients with osteochondromatosis.