Bram van Ginneken’s research while affiliated with Fraunhofer Institute for Digital Medicine and other places

Background Acute abdominal pain (AAP) constitutes 5–10% of all emergency department (ED) visits, with appendicitis being a prevalent AAP etiology often necessitating surgical intervention. The variability in AAP symptoms and causes, combined with the challenge of identifying appendicitis, complicate timely intervention. To estimate the risk of appendicitis, scoring systems such as the Alvarado score have been developed. However, diagnostic errors and delays remain common. Although various machine learning (ML) models have been proposed to enhance appendicitis detection, none have been seamlessly integrated into the ED workflows for AAP or are specifically designed to diagnose appendicitis as early as possible within the clinical decision-making process. To mimic daily clinical practice, this proof-of-concept study aims to develop ML models that support decision-making using comprehensive clinical data up to key decision points in the ED workflow to detect appendicitis in patients presenting with AAP. Methods Data from the Dutch triage system at the ED, vital signs, complete medical history and physical examination findings and routine laboratory test results were retrospectively extracted from 350 AAP patients presenting to the ED of a Dutch teaching hospital from 2016 to 2023. Two eXtreme Gradient Boosting ML models were developed to differentiate cases with appendicitis from other AAP causes: one model used all data up to and including physical examination, and the other was extended with routine laboratory test results. The performance of both models was evaluated on a validation set (n = 68) and compared to the Alvarado scoring system as well as three ED physicians in a reader study. Results The ML models achieved AUROCs of 0.919 without laboratory test results and 0.923 with the addition of laboratory test results. The Alvarado scoring system attained an AUROC of 0.824. ED physicians achieved AUROCs of 0.894, 0.826, and 0.791 without laboratory test results, increasing to AUROCs of 0.923, 0.892, and 0.859 with laboratory test results. Conclusions Both ML models demonstrated comparable high accuracy in predicting appendicitis in patients with AAP, outperforming the Alvarado scoring system. The ML models matched or surpassed ED physician performance in detecting appendicitis, with the largest potential performance gain observed in absence of laboratory test results. Integration could assist ED physicians in early and accurate diagnosis of appendicitis. Graphical abstract

Background : Acute abdominal pain (AAP) constitutes 5-10% of all emergency department (ED) visits, with appendicitis being a prevalent AAP etiology often necessitating surgical intervention. The variability in AAP symptoms and causes, combined with the challenge of identifying appendicitis, complicate timely intervention. To estimate the risk of appendicitis, scoring systems such as the Alvarado score have been developed. However, diagnostic errors and delays remain common. Although various machine learning (ML) models have been proposed to enhance appendicitis detection, none have been seamlessly integrated into the ED workflows for AAP or are specifically designed to diagnose appendicitis as early as possible within the clinical decision-making process. To mimic daily clinical practice, this proof-of-concept study aims to develop ML models that support decision-making using comprehensive clinical data up to key decision points in the ED workflow to detect appendicitis in patients presenting with AAP. Methods : Data from the Dutch triage system at the ED, vital signs, complete medical history and physical examination findings and routine laboratory test results were retrospectively extracted from 350 AAP patients presenting to the ED of a Dutch teaching hospital from 2016 to 2023. Two eXtreme Gradient Boosting ML models were developed to differentiate cases with appendicitis from other AAP causes: one model used all data up to and including physical examination, and the other was extended with routine laboratory test results. The performance of both models was evaluated on a validation set ( n =68) and compared to the Alvarado scoring system as well as three ED physicians in a reader study. Results: The ML models achieved AUROCs of 0.919 without laboratory test results and 0.923 with the addition of laboratory test results. The Alvarado scoring system attained an AUROC of 0.824. ED physicians achieved AUROCs of 0.894, 0.826, and 0.791 without laboratory test results, increasing to AUROCs of 0.923, 0.892, and 0.859 with laboratory test results. Conclusions: Both ML models demonstrated comparable high accuracy in predicting appendicitis in patients with AAP, outperforming the Alvarado scoring system. The ML models matched or surpassed ED physician performance in detecting appendicitis, with the largest potential performance gain observed in absence of laboratory test results. Integration could assist ED physicians in early and accurate diagnosis of appendicitis.

Objectives The main objective was to develop and evaluate an artificial intelligence (AI) model for tooth segmentation in magnetic resonance (MR) scans. Material and Methods MR scans of 20 patients performed with a commercial 64-channel head coil with a T1-weighted 3D-SPACE (Sampling Perfection with Application Optimized Contrasts using different flip angle Evolution) sequence were included. 16 datasets were used for model training and 4 for accuracy evaluation. Two clinicians segmented and annotated the teeth in each dataset. A segmentation model was trained using the nnU-Net framework. The manual reference tooth segmentation and the inferred tooth segmentation were superimposed and compared by computing precision, sensitivity, and Dice-Sørensen coefficient. Surface meshes were extracted from the segmentations, and the distances between points on each mesh and their closest counterparts on the other mesh were computed, of which the mean (average symmetric surface distance, ASSD) and 95th percentile (Hausdorff distance 95%, HD95) were reported. Results The model achieved an overall precision of 0.867, a sensitivity of 0.926, a Dice-Sørensen coefficient of 0.895, and a 95% Hausdorff distance of 0.91 mm. The model predictions were less accurate for datasets containing dental restorations due to image artefacts. Conclusions The current study developed an automated method for tooth segmentation in MR scans with moderate to high effectiveness for scans with respectively without artefacts.

Objectives The main objective was to develop and evaluate an artificial intelligence (AI) model for tooth segmentation in magnetic resonance (MR) scans. Material and Methods MR scans of 20 patients performed with a commercial 64-channel head coil with a T1-weighted 3D-SPACE (Sampling Perfection with Application Optimized Contrasts using different flip angle Evolution) sequence were included. 16 datasets were used for model training and 4 for accuracy evaluation. Two clinicians segmented and annotated the teeth in each dataset. A segmentation model was trained using the nnU-Net framework. The manual reference tooth segmentation and the inferred tooth segmentation were superimposed and compared by computing precision, sensitivity, and Dice-Sørensen coefficient. Surface meshes were extracted from the segmentations, and the distances between points on each mesh and their closest counterparts on the other mesh were computed, of which the mean (average symmetric surface distance, ASSD) and 95th percentile (Hausdorff distance 95%, HD95) were reported. Results The model achieved an overall precision of 0.867, a sensitivity of 0.926, a Dice-Sørensen coefficient of 0.895, and a 95% Hausdorff distance of 0.91 mm. The model predictions were less accurate for datasets containing dental restorations due to image artefacts. Conclusions The current study developed an automated method for tooth segmentation in MR scans with moderate to high effectiveness for scans with respectively without artefacts.

Introduction: Despite the notable progress in developing artificial intelligence-based tools for caries detection in bitewings, limited research has addressed the detection and staging of secondary caries. Therefore, we aimed to develop a convolutional neural network (CNN)-based algorithm for these purposes using a novel approach for determining lesion severity. Methods: We used a dataset from a Dutch dental practice-based research network containing 2,612 restored teeth in 413 bitewings from 383 patients aged 15-88 years and trained the Mask R-CNN architecture with a Swin Transformer backbone. Two-stage training fine-tuned caries detection accuracy and severity assessment. Annotations of caries around restorations were made by two evaluators and checked by two other experts. Aggregated accuracy metrics (mean ± standard deviation - SD) in detecting teeth with secondary caries were calculated considering two thresholds: detecting all lesions and dentine lesions. The correlation between the lesion severity scores obtained with the algorithm and the annotators' consensus was determined using the Pearson correlation coefficient and Bland-Altman plots. Results: Our refined algorithm showed high specificity in detecting all lesions (0.966 ± 0.025) and dentine lesions (0.964 ± 0.019). Sensitivity values were lower: 0.737 ± 0.079 for all lesions and 0.808 ± 0.083 for dentine lesions. The areas under ROC curves (SD) were 0.940 (0.025) for all lesions and 0.946 (0.023) for dentine lesions. The correlation coefficient for severity scores was 0.802. Conclusion: We developed an improved algorithm to support clinicians in detecting and staging secondary caries in bitewing, incorporating an innovative approach for annotation, considering the lesion severity as a continuous outcome.

Objectives The assessment of lumbar central canal stenosis (LCCS) is crucial for diagnosing and planning treatment for patients with low back pain and neurogenic pain. However, manual assessment methods are time-consuming, variable, and require axial MRIs. The aim of this study is to develop and validate an AI-based model that automatically classifies LCCS using sagittal T2-weighted MRIs. Methods A pre-existing 3D AI algorithm was utilized to segment the spinal canal and intervertebral discs (IVDs), enabling quantitative measurements at each IVD level. Four musculoskeletal radiologists graded 683 IVD levels from 186 LCCS patients using the 4-class Lee grading system. A second consensus reading was conducted by readers 1 and 2, which, along with automatic measurements, formed the training dataset for a multiclass (grade 0–3) and binary (grade 0–1 vs. 2–3) random forest classifier with tenfold cross-validation. Results The multiclass model achieved a Cohen’s weighted kappa of 0.86 (95% CI: 0.82–0.90), comparable to readers 3 and 4 with 0.85 (95% CI: 0.80–0.89) and 0.73 (95% CI: 0.68–0.79) respectively. The binary model demonstrated an AUC of 0.98 (95% CI: 0.97–0.99), sensitivity of 93% (95% CI: 91–96%), and specificity of 91% (95% CI: 87–95%). In comparison, readers 3 and 4 achieved a specificity of 98 and 99% and sensitivity of 74 and 54%, respectively. Conclusion Both the multiclass and binary models, while only using sagittal MR images, perform on par with experienced radiologists who also had access to axial sequences. This underscores the potential of this novel algorithm in enhancing diagnostic accuracy and efficiency in medical imaging. Key Points Question How can the classification of lumbar central canal stenosis (LCCS) be made more efficient? Findings Multiclass and binary AI models, using only sagittal MR images, performed on par with experienced radiologists who also had access to axial sequences. Clinical relevance Our AI algorithm accurately classifies LCCS from sagittal MRI, matching experienced radiologists. This study offers a promising tool for automated LCCS assessment from sagittal T2 MRI, potentially reducing the reliance on additional axial imaging .

Background To improve tuberculosis case-finding, rapid, non-sputum triage tests need to be developed according to the World Health Organization target product profile (TPP) (>90% sensitivity, >70% specificity). We prospectively evaluated and compared artificial intelligence–based, computer-aided detection software, CAD4TBv7, and C-reactive protein assay (CRP) as triage tests at health facilities in Lesotho and South Africa. Methods Adults (≥18 years) presenting with ≥1 of the 4 cardinal tuberculosis symptoms were consecutively recruited between February 2021 and April 2022. After informed consent, each participant underwent a digital chest X-ray for CAD4TBv7 and a CRP test. Participants provided 1 sputum sample for Xpert MTB/RIF Ultra and Xpert MTB/RIF and 1 for liquid culture. Additionally, an expert radiologist read the chest X-rays via teleradiology. For primary analysis, a composite microbiological reference standard (ie, positive culture or Xpert Ultra) was used. Results We enrolled 1392 participants, 48% were people with HIV and 24% had previously tuberculosis. The receiver operating characteristic curve for CAD4TBv7 and CRP showed an area under the curve of .87 (95% CI: .84–.91) and .80 (95% CI: .76–.84), respectively. At thresholds corresponding to 90% sensitivity, specificity was 68.2% (95% CI: 65.4–71.0%) and 38.2% (95% CI: 35.3–41.1%) for CAD4TBv7 and CRP, respectively. CAD4TBv7 detected tuberculosis as well as an expert radiologist. CAD4TBv7 almost met the TPP criteria for tuberculosis triage. Conclusions CAD4TBv7 is accurate as a triage test for patients with tuberculosis symptoms from areas with a high tuberculosis and HIV burden. The role of CRP in tuberculosis triage requires further research. Clinical Trials Registration Clinicaltrials.gov identifier: NCT04666311.

Background Hemoglobinopathies, the most common inherited blood disorder, are frequently underdiagnosed. Early identification of carriers is important for genetic counseling of couples at risk. The aim of this study was to develop and validate a novel machine learning model on a multicenter data set, covering a wide spectrum of hemoglobinopathies based on routine complete blood count (CBC) testing. Methods Hemoglobinopathy test results from 10 322 adults were extracted retrospectively from 8 Dutch laboratories. eXtreme Gradient Boosting (XGB) and logistic regression models were developed to differentiate negative from positive hemoglobinopathy cases, using 7 routine CBC parameters. External validation was conducted on a data set from an independent Dutch laboratory, with an additional external validation on a Spanish data set (n = 2629) specifically for differentiating thalassemia from iron deficiency anemia (IDA). Results The XGB and logistic regression models achieved an area under the receiver operating characteristic (AUROC) of 0.88 and 0.84, respectively, in distinguishing negative from positive hemoglobinopathy cases in the independent external validation set. Subclass analysis showed that the XGB model reached an AUROC of 0.97 for β-thalassemia, 0.98 for α0-thalassemia, 0.95 for homozygous α+-thalassemia, 0.78 for heterozygous α+-thalassemia, and 0.94 for the structural hemoglobin variants Hemoglobin C, Hemoglobin D, Hemoglobin E. Both models attained AUROCs of 0.95 in differentiating IDA from thalassemia. Conclusions Both the XGB and logistic regression model demonstrate high accuracy in predicting a broad range of hemoglobinopathies and are effective in differentiating hemoglobinopathies from IDA. Integration of these models into the laboratory information system facilitates automated hemoglobinopathy detection using routine CBC parameters.