The study aimed to develop a single-stage deep learning (DL) screening system for automated binary and multiclass grading of lumbar central stenosis (LCS), lateral recess stenosis (LRS), and lumbar foraminal stenosis (LFS). Consecutive inpatients who underwent lumbar MRI at our center were retrospectively reviewed for the internal dataset. Axial and sagittal lumbar MRI scans were collected. Based on a new MRI diagnostic criterion, all MRI studies were labeled by two spine specialists and calibrated by a third spine specialist to serve as reference standard. Furthermore, two spine clinicians labeled all MRI studies independently to compare interobserver reliability with the DL model. Samples were assigned into training, validation, and test sets at a proportion of 8:1:1. Additional patients from another center were enrolled as the external test dataset. A modified single-stage YOLOv5 network was designed for simultaneous detection of regions of interest (ROIs) and grading of LCS, LRS, and LFS. Quantitative evaluation metrics of exactitude and reliability for the model were computed. In total, 420 and 50 patients were enrolled in the internal and external datasets. High recalls of 97.4%-99.8% were achieved for ROI detection of lumbar spinal stenosis (LSS). The system revealed multigrade area under curve (AUC) values of 0.93-0.97 in the internal test set and 0.85-0.94 in the external test set for LCS, LRS, and LFS. In binary grading, the DL model achieved high sensitivities of 0.97 for LCS, 0.98 for LRS, and 0.96 for LFS, slightly better than those achieved by spine clinicians in the internal test set. In the external test set, the binary sensitivities were 0.98 for LCS, 0.96 for LRS, and 0.95 for LFS. For reliability assessment, the kappa coefficients between the DL model and reference standard were 0.92, 0.88, and 0.91 for LCS, LRS, and LFS, respectively, slightly higher than those evaluated by nonexpert spine clinicians. The authors designed a novel DL system that demonstrated promising performance, especially in sensitivity, for automated diagnosis and grading of different types of lumbar spinal stenosis using spine MRI. The reliability of the system was better than that of spine surgeons. The authors' system may serve as a triage tool for LSS to reduce misdiagnosis and optimize routine processes in clinical work.

We aimed to compare the diagnostic performance of physicians in the detection of arthroscopically confirmed meniscus and anterior cruciate ligament (ACL) tears on knee magnetic resonance imaging (MRI), with and without assistance from a deep learning (DL) model. We obtained preoperative MR images from 88 knees of patients who underwent arthroscopic meniscal repair, with or without ACL reconstruction. Ninety-eight MR images of knees without signs of meniscus or ACL tears were obtained from a publicly available database after matching on age and ACL status (normal or torn), resulting in a global dataset of 186 MRI examinations. The Keros^® (Incepto, Paris) DL algorithm, previously trained for the detection and characterization of meniscus and ACL tears, was used for MRI assessment. Magnetic resonance images were individually, and blindly annotated by three physicians and the DL algorithm. After three weeks, the three human raters repeated image assessment with model assistance, performed in a different order. The Keros^® algorithm achieved an area under the curve (AUC) of 0.96 (95% CI 0.93, 0.99), 0.91 (95% CI 0.85, 0.96), and 0.99 (95% CI 0.98, 0.997) in the detection of medial meniscus, lateral meniscus and ACL tears, respectively. With model assistance, physicians achieved higher sensitivity (91% vs. 83%, p = 0.04) and similar specificity (91% vs. 87%, p = 0.09) in the detection of medial meniscus tears. Regarding lateral meniscus tears, sensitivity and specificity were similar with/without model assistance. Regarding ACL tears, physicians achieved higher specificity when assisted by the algorithm (70% vs. 51%, p = 0.01) but similar sensitivity with/without model assistance (93% vs. 96%, p = 0.13). The current model consistently helped physicians in the detection of medial meniscus and ACL tears, notably when they were combined. Diagnostic study, Level III.

Cephalometric Landmark Detection (CLD), i.e. annotating interest points in lateral X-ray images, is the crucial first step of every orthodontic therapy. While CLD has immense potential for automation using Deep Learning methods, carefully crafted contemporary approaches using convolutional neural networks and heatmap prediction do not qualify for large-scale clinical application due to insufficient performance. We propose a novel approach using Vision Transformers (ViTs) with direct coordinate prediction, avoiding the memory-intensive heatmap prediction common in previous work. Through extensive ablation studies comparing our method against contemporary CNN architectures (ConvNext V2) and heatmap-based approaches (Segformer), we demonstrate that ViTs with coordinate prediction achieve superior performance with more than 2 mm improvement in mean radial error compared to state-of-the-art CLD methods. Our results show that while non-adapted CNN architectures perform poorly on the given task, contemporary approaches may be too tailored to specific datasets, failing to generalize to different and especially sparse datasets. We conclude that using general-purpose Vision Transformers with direct coordinate prediction shows great promise for future research on CLD and medical computer vision.

Early detection of lung cancer through low-dose CT lung cancer screening in a high-risk population has proven to reduce lung cancer-specific mortality. Nodule management plays a pivotal role in early detection and further diagnostic approaches. The European Society of Thoracic Imaging (ESTI) has established a nodule management recommendation to improve the handling of pulmonary nodules detected during screening. For solid nodules, the primary method for assessing the likelihood of malignancy is to monitor nodule growth using volumetry software. For subsolid nodules, the aggressiveness is determined by measuring the solid part. The ESTI-recommendation enhances existing protocols but puts a stronger focus on lesion aggressiveness. The main goals are to minimise the overall number of follow-up examinations while preventing the risk of a major stage shift and reducing the risk of overtreatment. KEY POINTS: Question Assessment of nodule growth and management according to guidelines is essential in lung cancer screening. Findings Assessment of nodule aggressiveness defines follow-up in lung cancer screening. Clinical relevance The ESTI nodule management recommendation aims to reduce follow-up examinations while preventing major stage shift and overtreatment.

The aim of this study was to develop and validate a fully automatic anatomical landmark localization and trajectory planning method for external ventricular drain (EVD) placement using CT or MRI. The authors used 125 preoperative CT and 137 contrast-enhanced T1-weighted MRI scans to generate 3D surface meshes of patients' skin and ventricular systems. Seven anatomical landmarks were manually annotated to train a neural network for automatic landmark localization. The model's accuracy was assessed by calculating the mean Euclidian distance of predicted landmarks to the ground truth. Kocher's point and EVD trajectories were automatically calculated with the foramen of Monro as the target. Performance was evaluated using Kakarla grades, as assessed by 3 clinicians. Interobserver agreement was measured with Pearson correlation, and scores were aggregated using majority voting. Ordinal linear regressions were used to assess whether modality or placement side had an effect on Kakarla grades. The impact of landmark localization error on the final EVD plan was also evaluated. The automated landmark localization model achieved a mean error of 4.0 mm (SD 2.6 mm). Trajectory planning generated a trajectory for all patients, with a Kakarla grade of 1 in 92.9% of cases. Statistical analyses indicated a strong interobserver agreement and no significant differences between modalities (CT vs MRI) or EVD placement sides. The location of Kocher's point and the target point were significantly correlated to nasion landmark localization error, with median drifts of 9.38 mm (95% CI 1.94-19.16 mm) and 3.91 mm (95% CI 0.18-26.76 mm) for Kocher's point and the target point, respectively. The presented method was efficient and robust for landmark localization and accurate EVD trajectory planning. The short processing time thereby also provides a base for use in emergency settings.

Oral cancer is a hazardous disease and a major cause of morbidity and mortality worldwide. The purpose of this study was to develop the deep convolutional neural networks (CNN)-based multiclass classification and object detection models for distinguishing and detection of oral carcinoma and sarcoma in contrast-enhanced CT images. This study included 3,259 slices of CT images of oral cancer cases from the cancer hospital and two regional hospitals from 2016 to 2020. Multiclass classification models were constructed using DenseNet-169, ResNet-50, EfficientNet-B0, ConvNeXt-Base, and ViT-Base-Patch16-224 to accurately differentiate between oral carcinoma and sarcoma. Additionally, multiclass object detection models, including Faster R-CNN, YOLOv8, and YOLOv11, were designed to autonomously identify and localize lesions by placing bounding boxes on CT images. Performance evaluation on a test dataset showed that the best classification model achieved an accuracy of 0.97, while the best detection models yielded a mean average precision (mAP) of 0.87. In conclusion, the CNN-based multiclass models have a great promise for accurately determining and distinguishing oral carcinoma and sarcoma in CT imaging, potentially enhancing early detection and informing treatment strategies.

The US healthcare system faces significant challenges, including clinician burnout, operational inefficiencies, and concerns about patient safety. Artificial intelligence (AI), particularly generative AI, has the potential to address these challenges, but its adoption, effectiveness, and barriers to implementation are not well understood. To evaluate the current state of AI adoption in US healthcare systems, assess successes and barriers to implementation during the early generative AI era. This cross-sectional survey was conducted in Fall 2024, and included 67 health systems members of the Scottsdale Institute, a collaborative of US non-profit healthcare organizations. Forty-three health systems completed the survey (64% response rate). Respondents provided data on the deployment status and perceived success of 37 AI use cases across 10 categories. The primary outcomes were the extent of AI use case development, piloting, or deployment, the degree of reported success for AI use cases, and the most significant barriers to adoption. Across the 43 responding health systems, AI adoption and perceptions of success varied significantly. Ambient Notes, a generative AI tool for clinical documentation, was the only use case with 100% of respondents reporting adoption activities, and 53% reported a high degree of success with using AI for Clinical Documentation. Imaging and radiology emerged as the most widely deployed clinical AI use case, with 90% of organizations reporting at least partial deployment, although successes with diagnostic use cases were limited. Similarly, many organizations have deployed AI for clinical risk stratification such as early sepsis detection, but only 38% report high success in this area. Immature AI tools were identified a significant barrier to adoption, cited by 77% of respondents, followed by financial concerns (47%) and regulatory uncertainty (40%). Ambient Notes is rapidly advancing in US healthcare systems and demonstrating early success. Other AI use cases show varying degrees of adoption and success, constrained by barriers such as immature AI tools, financial concerns, and regulatory uncertainty. Addressing these challenges through robust evaluations, shared strategies, and governance models will be essential to ensure effective integration and adoption of AI into healthcare practice.

Pancreatic ductal adenocarcinoma (PDAC) is the third leading cause of cancer-related deaths in the United States, largely due to its poor five-year survival rate and frequent late-stage diagnosis. A significant barrier to early detection even in high-risk cohorts is that the pancreas often appears morphologically normal during the pre-diagnostic phase. Yet, the disease can progress rapidly from subclinical stages to widespread metastasis, undermining the effectiveness of screening. Recently, artificial intelligence (AI) applied to cross-sectional imaging has shown significant potential in identifying subtle, early-stage changes in pancreatic tissue that are often imperceptible to the human eye. Moreover, AI-driven imaging also aids in the discovery of prognostic and predictive biomarkers, essential for personalized treatment planning. This article uniquely integrates a critical discussion on AI's role in detecting visually occult PDAC on pre-diagnostic imaging, addresses challenges of model generalizability, and emphasizes solutions like standardized datasets and clinical workflows. By focusing on both technical advancements and practical implementation, this article provides a forward-thinking conceptual framework that bridges current gaps in AI-driven PDAC research.

Early detection of breast cancer through mammography significantly improves survival rates. However, high false positive and false negative rates remain a challenge. Deep learning-based computer-aided diagnosis systems can assist in lesion detection, but their performance is often limited by the availability of labeled clinical data. This study systematically evaluated the effectiveness of transfer learning, image preprocessing techniques, and the latest You Only Look Once (YOLO) model (v9) for optimizing breast mass detection models on small proprietary datasets. We examined 133 mammography images containing masses and assessed various preprocessing strategies, including cropping and contrast enhancement. We further investigated the impact of transfer learning using the OPTIMAM Mammography Image Database (OMI-DB) compared with training on proprietary data. The performance of YOLOv9 was evaluated against YOLOv7 to determine improvements in detection accuracy. Pretraining on the OMI-DB dataset with cropped images significantly improved model performance, with YOLOv7 achieving a 13.9 % higher mean average precision (mAP) and 13.2 % higher F1-score compared to training only on proprietary data. Among the tested models and configurations, the best results were obtained using YOLOv9 pretrained OMI-DB and fine-tuned with cropped proprietary images, yielding an mAP of 73.3 % ± 16.7 % and an F1-score of 76.0 % ± 13.4 %, under this condition, YOLOv9 outperformed YOLOv7 by 8.1 % in mAP and 9.2 % in F1-score. This study provides a systematic evaluation of transfer learning and preprocessing techniques for breast mass detection in small datasets. Our results demonstrating that YOLOv9 with OMI-DB pretraining significantly enhances the performance of breast mass detection models while reducing training time, providing a valuable guideline for optimizing deep learning models in data-limited clinical applications.

In this paper, we introduce the Visual Query-based task of Video Clip Localization (VQ-VCL) for medical video understanding. Specifically, we aim to retrieve a video clip containing frames similar to a given exemplar frame from a given input video. To solve the task, we propose a novel visual query-based video clip localization model called TIER-LOC. TIER-LOC is designed to improve video clip retrieval, especially in fine-grained videos by extracting features from different levels, i.e., coarse to fine-grained, referred to as TIERS. The aim is to utilize multi-Tier features for detecting subtle differences, and adapting to scale or resolution variations, leading to improved video-clip retrieval. TIER-LOC has three main components: (1) a Multi-Tier Spatio-Temporal Transformer to fuse spatio-temporal features extracted from multiple Tiers of video frames with features from multiple Tiers of the visual query enabling better video understanding. (2) a Multi-Tier, Dual Anchor Contrastive Loss to deal with real-world annotation noise which can be notable at event boundaries and in videos featuring highly similar objects. (3) a Temporal Uncertainty-Aware Localization Loss designed to reduce the model sensitivity to imprecise event boundary. This is achieved by relaxing hard boundary constraints thus allowing the model to learn underlying class patterns and not be influenced by individual noisy samples. To demonstrate the efficacy of TIER-LOC, we evaluate it on two ultrasound video datasets and an open-source egocentric video dataset. First, we develop a sonographer workflow assistive task model to detect standard-frame clips in fetal ultrasound heart sweeps. Second, we assess our model's performance in retrieving standard-frame clips for detecting fetal anomalies in routine ultrasound scans, using the large-scale PULSE dataset. Lastly, we test our model's performance on an open-source computer vision video dataset by creating a VQ-VCL fine-grained video dataset based on the Ego4D dataset. Our model outperforms the best-performing state-of-the-art model by 7%, 4%, and 4% on the three video datasets, respectively.