Junior Radiologists Research Articles

Deep Learning for Distinguishing Mucinous Breast Carcinoma From Fibroadenoma on Ultrasound

PurposeMucinous breast carcinoma (MBC) tends to be misdiagnosed as fibroadenomas (FA) due to its benign imaging characteristics. We aimed to develop a deep learning (DL) model to differentiate MBC and FA based on ultrasound (US) images. The model could contribute to the diagnosis of MBC for radiologists. MethodsIn this retrospective study, 884 eligible patients (700 FA patients and 184 MBC patients) with 2257 US images were enrolled. The images were randomly divided into a training set (n = 1805 images) and a test set (n = 452 images) in a ratio of 8:2. First, we used the training set to establish DL model, DL+ age-cutoff model and DL+ age-tree model. Then, we compared the diagnostic performance of three models to get the optimal model. Finally, we evaluated the diagnostic performance of radiologists (4 junior and 4 senior radiologists) with and without the assistance of the optimal model in the test set. ResultsThe DL+ age-tree model yielded higher areas under the receiver operating characteristic curve (AUC) than DL model and DL+ age-cutoff model (0.945 vs. 0.835, P < .001; 0.945 vs. 0.931, P < .001, respectively). With the assistance of DL+ age-tree model, both junior and senior radiologists’ AUC had significant improvement (0.746-0.818, P = .010, 0.827-0.860, P = .005, respectively). ConclusionsThe DL+ age-tree model based on US images and age showed excellent performance in the differentiation of MBC and FA. Moreover, it can effectively improve the performance of radiologists with different degrees of experience that may contribute to reducing the misdiagnosis of MBC.

Automatic detection of adenoid hypertrophy on lateral nasopharyngeal radiographs of children based on deep learning.

Adenoid hypertrophy is a prevalent cause of upper airway obstruction in children, potentially leading to various otolaryngological complications and even systemic sequelae. The lateral nasopharyngeal radiograph is routinely employed for the diagnosis of adenoid hypertrophy. This study aimed to evaluate the accuracy and reliability of deep learning, using lateral nasopharyngeal radiographs, for the diagnosis of adenoid hypertrophy in pediatric patients. In the retrospective study, the lateral nasopharyngeal X-ray images were collected from children receiving therapy in the Children's Hospital of Soochow University, the 983th Hospital of Joint Logistics Support Forces of Chinese PLA and the Suzhou Wujiang District Children's Hospital from January 2023 to November 2023. Five deep learning models, i.e., AlexNet, VGG16, Inception v3, ResNet50 and DenseNet121, were used for model training and validation. Receiver operating characteristic (ROC) curve analyses were used to evaluate the performance of each model. The best algorithm was compared with interpretations from three radiologists on 208 images in the internal validation group. The lateral nasopharyngeal X-ray images were collected from 1,188 children, including 705 males (59.3%) and 483 females (40.7%), aged 8 months to 13 years, with a mean age of 5.57±2.66 years. Among the five deep learning models, DenseNet-121 performed the best, with area under the curve (AUC) values of 0.892 and 0.872, with accuracy of 0.895 and 0.878, sensitivity of 0.870 and 0.838, and specificity of 0.913 and 0.906 in the internal and external validation groups, respectively. The diagnostic performance of DenseNet-121 was higher than that of the junior and mid-level radiologists (0.892 vs. 0.836, 0.892 vs. 0.869), close to the senior radiologist (0.892 vs. 0.901). However, Delong's test revealed no significant difference between DenseNet121 and each radiologist in the validation group (P=0.24, P=0.52, P=0.79). All the five deep learning models in the study showed good performance for the diagnosis of adenoid hypertrophy, with DenseNet121 being the best, which was clinically relevant for the automatic identification of adenoid hypertrophy.

Deep learning-based detection of primary bone tumors around the knee joint on radiographs: a multicenter study.

Most primary bone tumors are often found in the bone around the knee joint. However, the detection of primary bone tumors on radiographs can be challenging for the inexperienced or junior radiologist. This study aimed to develop a deep learning (DL) model for the detection of primary bone tumors around the knee joint on radiographs. From four tertiary referral centers, we recruited 687 patients diagnosed with bone tumors (including osteosarcoma, chondrosarcoma, giant cell tumor of bone, bone cyst, enchondroma, fibrous dysplasia, etc.; 417 males, 270 females; mean age 22.8±13.2 years) by postoperative pathology or clinical imaging/follow-up, and 1,988 participants with normal bone radiographs (1,152 males, 836 females; mean age 27.9±12.2 years). The dataset was split into a training set for model development, an internal independent and an external test set for model validation. The trained model located bone tumor lesions and then detected tumor patients. Receiver operating characteristic curves and Cohen's kappa coefficient were used for evaluating detection performance. We compared the model's detection performance with that of two junior radiologists in the internal test set using permutation tests. The DL model correctly localized 94.5% and 92.9% bone tumors on radiographs in the internal and external test set, respectively. An accuracy of 0.964/0.920, and an area under the receiver operating characteristic curve (AUC) of 0.981/0.990 in DL detection of bone tumor patients were for the internal and external test set, respectively. Cohen's kappa coefficient of the model in the internal test set was significantly higher than that of the two junior radiologists with 4 and 3 years of experience in musculoskeletal radiology (Model vs. Reader A, 0.927 vs. 0.777, P<0.001; Model vs. Reader B, 0.927 vs. 0.841, P=0.033). The DL model achieved good performance in detecting primary bone tumors around the knee joint. This model had better performance than those of junior radiologists, indicating the potential for the detection of bone tumors on radiographs.

Development of deep learning model for diagnosing muscle-invasive bladder cancer on MRI with vision transformer

Rationale and objectivesTo develop and validate a deep learning (DL) model to automatically diagnose muscle-invasive bladder cancer (MIBC) on MRI with Vision Transformer (ViT). Materials and methodsThis multicenter retrospective study included patients with BC who reported to two institutions between January 2016 and June 2020 (training dataset) and a third institution between May 2017 and May 2022 (test dataset). The diagnostic model for MIBC and the segmentation model for BC on MRI were developed using the training dataset with 5-fold cross-validation. ViT- and convolutional neural network (CNN)-based diagnostic models were developed and compared for diagnostic performance using the area under the curve (AUC). The performance of the diagnostic model with manual and auto-generated regions of interest (ROImanual and ROIauto, respectively) was validated on the test dataset and compared to that of radiologists (three senior and three junior radiologists) using Vesical Imaging Reporting and Data System scoring. ResultsThe training and test datasets included 170 and 53 patients, respectively. Mean AUC of the top 10 ViT-based models with 5-fold cross-validation outperformed those of the CNN-based models (0.831 ± 0.003 vs. 0.713 ± 0.007–0.812 ± 0.006, p < .001). The diagnostic model with ROImanual achieved AUC of 0.872 (95 % CI: 0.777, 0.968), which was comparable to that of junior radiologists (AUC = 0.862, 0.873, and 0.930). Semi-automated diagnosis with the diagnostic model with ROIauto achieved AUC of 0.815 (95 % CI: 0.696, 0.935). ConclusionThe DL model effectively diagnosed MIBC. The ViT-based model outperformed CNN-based models, highlighting its utility in medical image analysis.

Artificial intelligence improves the diagnosis of human leukocyte antigen (HLA)-B27-negative axial spondyloarthritis based on multi-sequence magnetic resonance imaging and clinical features.

Axial spondyloarthritis (axSpA) is frequently diagnosed late, particularly in human leukocyte antigen (HLA)-B27-negative patients, resulting in a missed opportunity for optimal treatment. This study aimed to develop an artificial intelligence (AI) tool, termed NegSpA-AI, using sacroiliac joint (SIJ) magnetic resonance imaging (MRI) and clinical SpA features to improve the diagnosis of axSpA in HLA-B27-negative patients. We retrospectively included 454 HLA-B27-negative patients with rheumatologist-diagnosed axSpA or other diseases (non-axSpA) from the Third Affiliated Hospital of Southern Medical University and Nanhai Hospital between January 2010 and August 2021. They were divided into a training set (n=328) for 5-fold cross-validation, an internal test set (n=72), and an independent external test set (n=54). To construct a prospective test set, we further enrolled 87 patients between September 2021 and August 2023 from the Third Affiliated Hospital of Southern Medical University. MRI techniques employed included T1-weighted (T1W), T2-weighted (T2W), and fat-suppressed (FS) sequences. We developed NegSpA-AI using a deep learning (DL) network to differentiate between axSpA and non-axSpA at admission. Furthermore, we conducted a reader study involving 4 radiologists and 2 rheumatologists to evaluate and compare the performance of independent and AI-assisted clinicians. NegSpA-AI demonstrated superior performance compared to the independent junior rheumatologist (≤5 years of experience), achieving areas under the curve (AUCs) of 0.878 [95% confidence interval (CI): 0.786-0.971], 0.870 (95% CI: 0.771-0.970), and 0.815 (95% CI: 0.714-0.915) on the internal, external, and prospective test sets, respectively. The assistance of NegSpA-AI promoted discriminating accuracy, sensitivity, and specificity of independent junior radiologists by 7.4-11.5%, 1.0-13.3%, and 7.4-20.6% across the 3 test sets (all P<0.05). On the prospective test set, AI assistance also improved the diagnostic accuracy, sensitivity, and specificity of independent junior rheumatologists by 7.7%, 7.7%, and 6.9%, respectively (all P<0.01). The proposed NegSpA-AI effectively improves radiologists' interpretations of SIJ MRI and rheumatologists' diagnoses of HLA-B27-negative axSpA.

Diagnostic Accuracy of Ultra-Low Dose CT Compared to Standard Dose CT for Identification of Fresh Rib Fractures by Deep Learning Algorithm.

The present study aimed to evaluate the diagnostic accuracy of ultra-low dose computed tomography (ULD-CT) compared to standard dose computed tomography (SD-CT) in discerning recent rib fractures using a deep learning algorithm detection of rib fractures (DLADRF). A total of 158 patients undergoing forensic diagnosis for rib fractures were included in this study: 50 underwent SD-CT, and 108 were assessed using ULD-CT. Junior and senior radiologists independently evaluated the images to identify and characterize the rib fractures. The sensitivity of rib fracture diagnosis by radiologists and radiologist + DLADRF was better using SD-CT than ULD-CT. However, the diagnosis sensitivity of DLADRF using ULD-CT alone was slightly more than SD-CT. Nonetheless, no substantial differences were observed in specificity, positive predictive value, and negative predictive value between SD-CT and ULD-CT by the same radiologist, radiologist + DLADRF, and DLADRF (P > 0.05). The area under the curve (AUC) of receiver operating characteristic indicated that senior radiologist + DLADRF was significantly better than senior and junior radiologists, junior radiologists + DLADRF, and DLADRF alone using SD-CT or ULD-CT (all P < 0.05). Also, junior radiologists + DLADRF was better with ULD-CT than senior and junior radiologists (P < 0.05). The AUC of the rib fracture diagnosed by senior radiologists did not differ from DLADRF using ULD-CT. Also, no significant differences were observed between junior + AI and senior and between junior and DLADRF using SD-CT. DLADRF enhanced the diagnostic performance of radiologists in detecting recent rib fractures. The diagnostic outcomes between SD-CT and ULD-CT across radiologists' experience and DLADRF did not differ significantly.

Prospective and multi-reader evaluation of deep learning reconstruction-based accelerated rectal MRI: image quality, diagnostic performance, and reading time.

To evaluate deep learning reconstruction (DLR)-based accelerated rectal magnetic resonance imaging (MRI) compared with standard MRI. Patients with biopsy-confirmed rectal adenocarcinoma between November/2022 and May/2023 in a single centre were prospectively enrolled for an intra-individual comparison between standard fast spin-echo (FSEstandard) and DLR-based FSE (FSEDL) sequences. Quantitative and qualitative image quality metrics of the pre-therapeutic MRIs were evaluated in all patients; diagnostic performance and evaluating time for T-staging, N-staging, extramural vascular invasion (EMVI), and mesorectal fascia (MRF) status was further analysed in patients undergoing curative surgery, with histopathologic results as the diagnostic gold standard. A total of 117 patients were enrolled, with 60 patients undergoing curative surgery. FSEDL reduced the acquisition time by 65% than FSEstandard. FSEDL exhibited higher signal-to-noise ratios, contrast-to-noise ratio, and subjective scores (noise, tumour margin clarity, visualisation of bowel wall layering and MRF, overall image quality, and diagnostic confidence) than FSEstandard (p < 0.001). Reduced artefacts were observed in FSEDL for patients without spasmolytics (p < 0.05). FSEDL provided higher T-staging accuracy by junior readers than FSEstandard (reader 1, 58.33% vs 70.00%, p = 0.016; reader 3, 60.00% vs 76.67%, p = 0.021), with similar N-staging, EMVI, and MRF performance. No significant difference was observed for senior readers. FSEDL exhibited shorter diagnostic time in all readers' T-staging and overall evaluation, and junior readers' EMVI and MRF (p < 0.05). FSEDL provided improved image quality, reading time, and junior radiologists' T-staging accuracy than FSEstandard, while reducing the acquisition time by 65%. DLR is clinically applicable for rectal MRI, providing improved image quality with shorter scanning time, which may ease the examination burden. It is beneficial for diagnostic optimisation in improving junior radiologists' T-staging accuracy and reading time. The rising incidence of rectal cancer has demanded enhanced efficiency and quality in imaging examinations. FSEDL demonstrated superior image quality and had a 65% reduced acquisition time. FSEDL can improve the diagnostic accuracy of T-staging and reduce the reading time for assessing rectal cancer.

Fully and Weakly Supervised Deep Learning for Meniscal Injury Classification, and Location Based on MRI.

Meniscal injury is a common cause of knee joint pain and a precursor to knee osteoarthritis (KOA). The purpose of this study is to develop an automatic pipeline for meniscal injury classification and localization using fully and weakly supervised networks based on MRI images. In this retrospective study, data were from the osteoarthritis initiative (OAI). The MR images were reconstructed using a sagittal intermediate-weighted fat-suppressed turbo spin-echo sequence. (1) We used 130 knees from the OAI to develop the LGSA-UNet model which fuses the features of adjacent slices and adjusts the blocks in Siam to enable the central slice to obtain rich contextual information. (2) One thousand seven hundred and fifty-six knees from the OAI were included to establish segmentation and classification models. The segmentation model achieved a DICE coefficient ranging from 0.84 to 0.93. The AUC values ranged from 0.85 to 0.95 in the binary models. The accuracy for the three types of menisci (normal, tear, and maceration) ranged from 0.60 to 0.88. Furthermore, 206 knees from the orthopedic hospital were used as an external validation data set to evaluate the performance of the model. The segmentation and classification models still performed well on the external validation set. To compare the diagnostic performances between the deep learning (DL) models and radiologists, the external validation sets were sent to two radiologists. The binary classification model outperformed the diagnostic performance of the junior radiologist (0.82-0.87 versus 0.74-0.88). This study highlights the potential of DL in knee meniscus segmentation and injury classification which can help improve diagnostic efficiency.

Research on automated rapid diagnosis of distal radius fractures using X-ray images

The primary diagnostic method for distal radius fractures remains conventional X-ray imaging. In emergency situations, rapid assessment and decision-making are often required from non-orthopedic physicians or junior radiologists in the emergency room. Therefore, an accurate and efficient auxiliary diagnostic technology for distal radius fractures is of great significance. This paper combines deep learning with image analysis techniques to propose an effective classification method for types of distal radius fractures. Firstly, an extended U-Net three-level cascade segmentation network is used for precise segmentation of the most critical joint surface and non-joint surface areas for fracture identification. The edge is further optimized using the central pixel image block classification method. Then, fracture identification is performed separately on the joint surface and non-joint surface areas. Finally, based on the classification results of the two types of images, a comprehensive judgment is made to determine normal or ABC fracture types. The experimental X-ray data comes from three tertiary hospitals, with a total of 12,000 images in the training set, 3,000 of each type, and 1,200 images in the test set, 300 of each type. To further demonstrate the generalization performance of the classification experiment, an additional 500 images were added for fracture classification testing, coming from collaborative medical institutions in the United States and Germany. All images were annotated by orthopedic medical experts with more than ten years of experience. The accuracy rates for normal, type A, type B, and type C fractures in the test set were 0.99, 0.92, 0.91, and 0.82, respectively. For orthopedic medical experts, the average recognition accuracy rates were 0.98, 0.90, 0.87, and 0.81. The proposed automatic recognition method overall performs better than experts and can be used for preliminary auxiliary diagnosis of distal radius fractures without expert participation.

Research on automated rapid diagnosis of distal radius fractures using X-ray images

The primary diagnostic method for distal radius fractures remains conventional X-ray imaging. In emergency situations, rapid assessment and decision-making are often required from non-orthopedic physicians or junior radiologists in the emergency room. Therefore, an accurate and efficient auxiliary diagnostic technology for distal radius fractures is of great significance. This paper combines deep learning with image analysis techniques to propose an effective classification method for types of distal radius fractures. Firstly, an extended U-Net three-level cascade segmentation network is used for precise segmentation of the most critical joint surface and non-joint surface areas for fracture identification. The edge is further optimized using the central pixel image block classification method. Then, fracture identification is performed separately on the joint surface and non-joint surface areas. Finally, based on the classification results of the two types of images, a comprehensive judgment is made to determine normal or ABC fracture types. The experimental X-ray data comes from three tertiary hospitals, with a total of 12,000 images in the training set, 3,000 of each type, and 1,200 images in the test set, 300 of each type. To further demonstrate the generalization performance of the classification experiment, an additional 500 images were added for fracture classification testing, coming from collaborative medical institutions in the United States and Germany. All images were annotated by orthopedic medical experts with more than ten years of experience. The accuracy rates for normal, type A, type B, and type C fractures in the test set were 0.99, 0.92, 0.91, and 0.82, respectively. For orthopedic medical experts, the average recognition accuracy rates were 0.98, 0.90, 0.87, and 0.81. The proposed automatic recognition method overall performs better than experts and can be used for preliminary auxiliary diagnosis of distal radius fractures without expert participation.

The clinical value of artificial intelligence in assisting junior radiologists in thyroid ultrasound: a multicenter prospective study from real clinical practice

BackgroundThis study is to propose a clinically applicable 2-echelon (2e) diagnostic criteria for the analysis of thyroid nodules such that low-risk nodules are screened off while only suspicious or indeterminate ones are further examined by histopathology, and to explore whether artificial intelligence (AI) can provide precise assistance for clinical decision-making in the real-world prospective scenario.MethodsIn this prospective study, we enrolled 1036 patients with a total of 2296 thyroid nodules from three medical centers. The diagnostic performance of the AI system, radiologists with different levels of experience, and AI-assisted radiologists with different levels of experience in diagnosing thyroid nodules were evaluated against our proposed 2e diagnostic criteria, with the first being an arbitration committee consisting of 3 senior specialists and the second being cyto- or histopathology.ResultsAccording to the 2e diagnostic criteria, 1543 nodules were classified by the arbitration committee, and the benign and malignant nature of 753 nodules was determined by pathological examinations. Taking pathological results as the evaluation standard, the sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) of the AI systems were 0.826, 0.815, 0.821, and 0.821. For those cases where diagnosis by the Arbitration Committee were taken as the evaluation standard, the sensitivity, specificity, accuracy, and AUC of the AI system were 0.946, 0.966, 0.964, and 0.956. Taking the global 2e diagnostic criteria as the gold standard, the sensitivity, specificity, accuracy, and AUC of the AI system were 0.868, 0.934, 0.917, and 0.901, respectively. Under different criteria, AI was comparable to the diagnostic performance of senior radiologists and outperformed junior radiologists (all P < 0.05). Furthermore, AI assistance significantly improved the performance of junior radiologists in the diagnosis of thyroid nodules, and their diagnostic performance was comparable to that of senior radiologists when pathological results were taken as the gold standard (all p > 0.05).ConclusionsThe proposed 2e diagnostic criteria are consistent with real-world clinical evaluations and affirm the applicability of the AI system. Under the 2e criteria, the diagnostic performance of the AI system is comparable to that of senior radiologists and significantly improves the diagnostic capabilities of junior radiologists. This has the potential to reduce unnecessary invasive diagnostic procedures in real-world clinical practice.

Artificial intelligence-based graded training of pulmonary nodules for junior radiology residents and medical imaging students

BackgroundTo evaluate the efficiency of artificial intelligence (AI)-assisted diagnosis system in the pulmonary nodule detection and diagnosis training of junior radiology residents and medical imaging students.MethodsThe participants were divided into three groups. Medical imaging students of Grade 2020 in the Jinzhou Medical University were randomly divided into Groups 1 and 2; Group 3 comprised junior radiology residents. Group 1 used the traditional case-based teaching mode; Groups 2 and 3 used the ‘AI intelligent assisted diagnosis system’ teaching mode. All participants performed localisation, grading and qualitative diagnosed of 1,057 lung nodules in 420 cases for seven rounds of testing after training. The sensitivity and number of false positive nodules in different densities (solid, pure ground glass, mixed ground glass and calcification), sizes (less than 5 mm, 5–10 mm and over 10 mm) and positions (subpleural, peripheral and central) of the pulmonary nodules in the three groups were detected. The pathological results and diagnostic opinions of radiologists formed the criteria. The detection rate, diagnostic compliance rate, false positive number/case, and kappa scores of the three groups were compared.ResultsThere was no statistical difference in baseline test scores between Groups 1 and 2, and there were statistical differences with Group 3 (P = 0.036 and 0.011). The detection rate of solid, pure ground glass and calcified nodules; small-, medium-, and large-diameter nodules; and peripheral nodules were significantly different among the three groups (P<0.05). After seven rounds of training, the diagnostic compliance rate increased in all three groups, with the largest increase in Group 2. The average kappa score increased from 0.508 to 0.704. The average kappa score for Rounds 1–4 and 5–7 were 0.595 and 0.714, respectively. The average kappa scores of Groups 1,2 and 3 increased from 0.478 to 0.658, 0.417 to 0.757, and 0.638 to 0.791, respectively.ConclusionThe AI assisted diagnosis system is a valuable tool for training junior radiology residents and medical imaging students to perform pulmonary nodules detection and diagnosis.

Artificial Intelligence Application in Skull Bone Fracture with Segmentation Approach.

This study aims to evaluate an AI model designed to automatically classify skull fractures and visualize segmentation on emergent CT scans. The model's goal is to boost diagnostic accuracy, alleviate radiologists' workload, and hasten diagnosis, thereby enhancing patient outcomes. Unique to this research, both pediatric and post-operative patients were not excluded, and diagnostic durations were analyzed. Our testing dataset for theobserver studies involved 671 patients, with a mean age of 58.88 years and fairly balanced gender representation. Model 1 of our AI algorithm, trained with 1499 fracture-positive cases, showed a sensitivity of 0.94 and specificity of 0.87, with a DICE score of 0.65. Implementing post-processing rules (specifically Rule B) improved the model's performance, resulting in a sensitivity of 0.94, specificity of 0.99, and a DICE score of 0.63. AI-assisted diagnosis resulted in significantly enhanced performance for all participants, with sensitivity almost doubling for junior radiology residents and other specialists. Additionally, diagnostic durations were significantly reduced (p < 0.01) with AI assistance across all participant categories. Our skull fracture detection model, employing a segmentation approach, demonstrated high performance, enhancing diagnostic accuracy and efficiency for radiologists and clinical physicians. This underlines the potential of AI integration in medical imaging analysis to improve patient care.

Deep learning-based differentiation of peripheral high-flow and low-flow vascular malformations in T2-weighted short tau inversion recovery MRI.

Differentiation of high-flow from low-flow vascular malformations (VMs) is crucial for therapeutic management of this orphan disease. A convolutional neural network (CNN) was evaluated for differentiation of peripheral vascular malformations (VMs) on T2-weighted short tau inversion recovery (STIR) MRI. 527 MRIs (386 low-flow and 141 high-flow VMs) were randomly divided into training, validation and test set for this single-center study. 1) Results of the CNN's diagnostic performance were compared with that of two expert and four junior radiologists. 2) The influence of CNN's prediction on the radiologists' performance and diagnostic certainty was evaluated. 3) Junior radiologists' performance after self-training was compared with that of the CNN. Compared with the expert radiologists the CNN achieved similar accuracy (92% vs. 97%, p = 0.11), sensitivity (80% vs. 93%, p = 0.16) and specificity (97% vs. 100%, p = 0.50). In comparison to the junior radiologists, the CNN had a higher specificity and accuracy (97% vs. 80%, p < 0.001; 92% vs. 77%, p < 0.001). CNN assistance had no significant influence on their diagnostic performance and certainty. After self-training, the junior radiologists' specificity and accuracy improved and were comparable to that of the CNN. Diagnostic performance of the CNN for differentiating high-flow from low-flow VM was comparable to that of expert radiologists. CNN did not significantly improve the simulated daily practice of junior radiologists, self-training was more effective.

HE-Mind: A model for automatically predicting hematoma expansion after spontaneous intracerebral hemorrhage

PurposeTo develop and validate an end-to-end model for automatically predicting hematoma expansion (HE) after spontaneous intracerebral hemorrhage (sICH) using a novel deep learning framework. MethodsThis multicenter retrospective study collected cranial noncontrast computed tomography (NCCT) images of 490 patients with sICH at admission for model training (n = 236), internal testing (n = 60), and external testing (n = 194). A HE-Mind model was designed to predict HE, which consists of a densely connected U-net for segmentation process, a multi-instance learning strategy for resolving label ambiguity and a Siamese network for classification process. Two radiomics models based on support vector machine or logistic regression and two deep learning models based on residual network or Swin transformer were developed for performance comparison. Reader experiments including physician diagnosis mode and artificial intelligence mode were conducted for efficiency comparison. ResultsThe HE-Mind model showed better performance compared to the comparative models in predicting HE, with areas under the curve of 0.849 and 0.809 in the internal and external test sets respectively. With the assistance of the HE-Mind model, the predictive accuracy and work efficiency of the emergency physician, junior radiologist, and senior radiologist were significantly improved, with accuracies of 0.768, 0.789, and 0.809 respectively, and reporting times of 7.26 s, 5.08 s, and 3.99 s respectively. ConclusionsThe HE-Mind model could rapidly and automatically process the NCCT data and predict HE after sICH within three seconds, indicating its potential to assist physicians in the clinical diagnosis workflow of HE.

Deep learning driven diagnosis of malignant soft tissue tumors based on dual-modal ultrasound images and clinical indexes

BackgroundSoft tissue tumors (STTs) are benign or malignant superficial neoplasms arising from soft tissues throughout the body with versatile pathological types. Although Ultrasonography (US) is one of the most common imaging tools to diagnose malignant STTs, it still has several drawbacks in STT diagnosis that need improving.ObjectivesThe study aims to establish this deep learning (DL) driven Artificial intelligence (AI) system for predicting malignant STTs based on US images and clinical indexes of the patients.MethodsWe retrospectively enrolled 271 malignant and 462 benign masses to build the AI system using 5-fold validation. A prospective dataset of 44 malignant masses and 101 benign masses was used to validate the accuracy of system. A multi-data fusion convolutional neural network, named ultrasound clinical soft tissue tumor net (UC-STTNet), was developed to combine gray scale and color Doppler US images and clinic features for malignant STTs diagnosis. Six radiologists (R1-R6) with three experience levels were invited for reader study.ResultsThe AI system achieved an area under receiver operating curve (AUC) value of 0.89 in the retrospective dataset. The diagnostic performance of the AI system was higher than that of one of the senior radiologists (AUC of AI vs R2: 0.89 vs. 0.84, p=0.022) and all of the intermediate and junior radiologists (AUC of AI vs R3, R4, R5, R6: 0.89 vs 0.75, 0.81, 0.80, 0.63; p &amp;lt;0.01). The AI system also achieved an AUC of 0.85 in the prospective dataset. With the assistance of the system, the diagnostic performances and inter-observer agreement of the radiologists was improved (AUC of R3, R5, R6: 0.75 to 0.83, 0.80 to 0.85, 0.63 to 0.69; p&amp;lt;0.01).ConclusionThe AI system could be a useful tool in diagnosing malignant STTs, and could also help radiologists improve diagnostic performance.

A domain knowledge-based interpretable deep learning system for improving clinical breast ultrasound diagnosis.

Though deep learning has consistently demonstrated advantages in the automatic interpretation of breast ultrasound images, its black-box nature hinders potential interactions with radiologists, posing obstacles for clinical deployment. We proposed a domain knowledge-based interpretable deep learning system for improving breast cancer risk prediction via paired multimodal ultrasound images. The deep learning system was developed on 4320 multimodal breast ultrasound images of 1440 biopsy-confirmed lesions from 1348 prospectively enrolled patients across two hospitals between August 2019 and December 2022. The lesions were allocated to 70% training cohort, 10% validation cohort, and 20% test cohort based on case recruitment date. Here, we show that the interpretable deep learning system can predict breast cancer risk as accurately as experienced radiologists, with an area under the receiver operating characteristic curve of 0.902 (95% confidence interval = 0.882 - 0.921), sensitivity of 75.2%, and specificity of 91.8% on the test cohort. With the aid of the deep learning system, particularly its inherent explainable features, junior radiologists tend to achieve better clinical outcomes, while senior radiologists experience increased confidence levels. Multimodal ultrasound images augmented with domain knowledge-based reasoning cues enable an effective human-machine collaboration at a high level of prediction performance. Such a clinically applicable deep learning system may be incorporated into future breast cancer screening and support assisted or second-read workflows.

Ultrasound image-based nomogram combining clinical, radiomics, and deep transfer learning features for automatic classification of ovarian masses according to O-RADS.

Accurate and rapid discrimination between benign and malignant ovarian masses is crucial for optimal patient management. This study aimed to establish an ultrasound image-based nomogram combining clinical, radiomics, and deep transfer learning features to automatically classify the ovarian masses into low risk and intermediate-high risk of malignancy lesions according to the Ovarian- Adnexal Reporting and Data System (O-RADS). The ultrasound images of 1,080 patients with 1,080 ovarian masses were included. The training cohort consisting of 683 patients was collected at the South China Hospital of Shenzhen University, and the test cohort consisting of 397 patients was collected at the Shenzhen University General Hospital. The workflow included image segmentation, feature extraction, feature selection, and model construction. The pre-trained Resnet-101 model achieved the best performance. Among the different mono-modal features and fusion feature models, nomogram achieved the highest level of diagnostic performance (AUC: 0.930, accuracy: 84.9%, sensitivity: 93.5%, specificity: 81.7%, PPV: 65.4%, NPV: 97.1%, precision: 65.4%). The diagnostic indices of the nomogram were higher than those of junior radiologists, and the diagnostic indices of junior radiologists significantly improved with the assistance of the model. The calibration curves showed good agreement between the prediction of nomogram and actual classification of ovarian masses. The decision curve analysis showed that the nomogram was clinically useful. This model exhibited a satisfactory diagnostic performance compared to junior radiologists. It has the potential to improve the level of expertise of junior radiologists and provide a fast and effective method for ovarian cancer screening.

Artificial intelligence-based tools with automated segmentation and measurement on CT images to assist accurate and fast diagnosis in acute pancreatitis.

To develop an artificial intelligence (AI) tool with automated pancreas segmentation and measurement of pancreatic morphological information on CT images to assist improved and faster diagnosis in acute pancreatitis. This study retrospectively contained 1124 patients suspected for AP and received non-contrast and enhanced abdominal CT examination between September 2013 and September 2022. Patients were divided into training (N = 688), validation (N = 145), testing dataset [N = 291; N = 104 for normal pancreas, N = 98 for AP, N = 89 for AP complicated with PDAC (AP&PDAC)]. A model based on convolutional neural network (MSAnet) was developed. The pancreas segmentation and measurement were performed via eight open-source models and MSAnet based tools, and the efficacy was evaluated using dice similarity coefficient (DSC) and intersection over union (IoU). The DSC and IoU for patients with different ages were also compared. The outline of tumour and oedema in the AP and were segmented by clustering. The diagnostic efficacy for radiologists with or without the assistance of MSAnet tool in AP and AP&PDAC was evaluated using receiver operation curve and confusion matrix. Among all models, MSAnet based tool showed best performance on the training and validation dataset, and had high efficacy on testing dataset. The performance was age-affected. With assistance of the AI tool, the diagnosis time was significantly shortened by 26.8% and 32.7% for junior and senior radiologists, respectively. The area under curve (AUC) in diagnosis of AP was improved from 0.91 to 0.96 for junior radiologist and 0.98 to 0.99 for senior radiologist. In AP&PDAC diagnosis, AUC was increased from 0.85 to 0.92 for junior and 0.97 to 0.99 for senior. MSAnet based tools showed good pancreas segmentation and measurement performance, which help radiologists improve diagnosis efficacy and workflow in both AP and AP with PDAC conditions. This study developed an AI tool with automated pancreas segmentation and measurement and provided evidence for AI tool assistance in improving the workflow and accuracy of AP diagnosis.

A radiograph-based deep learning model improves radiologists’ performance for classification of histological types of primary bone tumors: A multicenter study

PurposeTo develop a deep learning (DL) model for classifying histological types of primary bone tumors (PBTs) using radiographs and evaluate its clinical utility in assisting radiologists. MethodsThis retrospective study included 878 patients with pathologically confirmed PBTs from two centers (638, 77, 80, and 83 for the training, validation, internal test, and external test sets, respectively). We classified PBTs into five categories by histological types: chondrogenic tumors, osteogenic tumors, osteoclastic giant cell-rich tumors, other mesenchymal tumors of bone, or other histological types of PBTs. A DL model combining radiographs and clinical features based on the EfficientNet-B3 was developed for five-category classification. The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity were calculated to evaluate model performance. The clinical utility of the model was evaluated in an observer study with four radiologists. ResultsThe combined model achieved a macro average AUC of 0.904/0.873, with an accuracy of 67.5 %/68.7 %, a macro average sensitivity of 66.9 %/57.2 %, and a macro average specificity of 92.1 %/91.6 % on the internal/external test set, respectively. Model-assisted analysis improved accuracy, interpretation time, and confidence for junior (50.6 % vs. 72.3 %, 53.07[s] vs. 18.55[s] and 3.10 vs. 3.73 on a 5-point Likert scale [P < 0.05 for each], respectively) and senior radiologists (68.7 % vs. 75.3 %, 32.50[s] vs. 21.42[s] and 4.19 vs. 4.37 [P < 0.05 for each], respectively). ConclusionThe combined DL model effectively classified histological types of PBTs and assisted radiologists in achieving better classification results than their independent visual assessment.