Deep learning models to predict primary open-angle glaucoma.

Glaucoma is a major cause of blindness and vision impairment worldwide, and visual field (VF) tests are essential for monitoring the conversion of glaucoma. While previous studies have primarily focused on using VF data at a single time point for glaucoma prediction, there has been limited exploration of longitudinal trajectories. Additionally, many deep learning techniques treat the time-to-glaucoma prediction as a binary classification problem (glaucoma Yes/No), resulting in the misclassification of some censored subjects into the nonglaucoma category and decreased power. To tackle these challenges, we propose and implement several deep-learning approaches that naturally incorporate temporal and spatial information from longitudinal VF data to predict time-to-glaucoma. When evaluated on the Ocular Hypertension Treatment Study (OHTS) dataset, our proposed convolutional neural network (CNN)-long short-term memory (LSTM) emerged as the top-performing model among all those examined. The implementation code can be found online (https://github.com/rivenzhou/VF_prediction).

A real-time vehicle safety system by concurrent object detection and head pose estimation via stereo vision.

A considerable number of vehicular accidents occur in low-millage zones like school streets, neighborhoods, and parking lots, among others. Therefore, the proposed work aims to provide a novel ADAS system to warn about dangerous scenarios by analyzing the driver's attention and the corresponding distances between the vehicle and the detected object on the road. This approach is made possible by concurrent Head Pose Estimation (HPE) and Object/Pedestrian Detection. Both approaches have shown independently their viable application in the automotive industry to decrease the number of vehicle collisions. The proposed system takes advantage of stereo vision characteristics for HPE by enabling the computation of the Euler Angles with a low average error for classifying the driver's attention on the road using neural networks. For Object Detection, stereo vision is used to detect the distance between the vehicle and the approaching object; this is made with a state-of-the-art algorithm known as YOLO-R and a fast template matching technique known as SoRA that provides lower processing times. The result is an ADAS system designed to ensure adequate braking time, considering the driver's attention on the road and the distances to objects.

Super-resolution landmark detection networks for medical images.

Craniomaxillofacial (CMF) and nasal landmark detection are fundamental components in computer-assisted surgery. Medical landmark detection method includes regression-based and heatmap-based methods, and heatmap-based methods are among the main methodology branches. The method relies on high-resolution (HR) features containing more location information to reduce the network error caused by sub-pixel location. Previous studies extracted HR patches around each landmark from downsampling images via object detection and subsequently input them into the network to obtain HR features. Complex multistage tasks affect accuracy. The network error caused by downsampling and upsampling operations during training, which interpolates low-resolution features to generate HR features or predicted heatmap, is still significant. We propose standard super-resolution landmark detection networks (SRLD-Net) and super-resolution UNet (SR-UNet) to reduce network error effectively. SRLD-Net used Pyramid pooling block, Pyramid fusion block and super-resolution fusion block to combine global prior knowledge and multi-scale local features, similarly, SR-UNet adopts Pyramid pooling block and super-resolution block. They can obviously improve representation learning ability of our proposed methods. Then the super-resolution upsampling layer is utilized to generate detail predicted heatmap. Our proposed networks were compared to state-of-the-art methods using the craniomaxillofacial, nasal, and mandibular molar datasets, demonstrating better performance. The mean errors of 18 CMF, 6 nasal and 14 mandibular landmarks are 1.39 ± 1.04, 1.31 ± 1.09, 2.01 ± 4.33 mm. These results indicate that the super-resolution methods have great potential in medical landmark detection tasks. This paper provides two effective heatmap-based landmark detection networks and the code is released in https://github.com/Runshi-Zhang/SRLD-Net.

The surgical role of C1 nerve root identification for the disconnection of the spinal dural arteriovenous fistula at the craniocervical junction: a single center restrospective experience.

This study aims to discuss the identification of the C1 nerve root as an effective surgical approach to successfully locate the shunting point of craniocervical junction spinal dural arteriovenous fistula (CCJ-SDAVF) intraoperatively. This study included all patients with CCJ-SDAVF who underwent surgical treatment using the far-lateral transcondylar approach at a single institution from January 2017 to June 2023. Data on patient demographics, clinical and angiographic characteristics of CCJ-SDAVF, surgical details, and treatment outcomes were collected. Follow-up assessments were conducted for all patients until December 31, 2023. The study included a total of 7 patients, comprising 5 men(71.4%) and 2 women (28.6%), with an average age of 57.6 years. Among them, 4 patients (57.1%) developed diffuse subarachnoid hemorrhage(SAH), while 2 patients (28.6%) experienced progressive cervical myelopathy. The shunting points of all CCJ-SDAVFs, which exhibited engorged veins, were identified next to the C1 root. Complete obliteration of CCJ-SDAVFs was successfully achieved in all patients, as confirmed by postoperative angiography one month later. No recurrent CCJ-SDAVFs were observed two years after the operation. Among the patients, 5 (71.4%) experienced good functional recovery, as indicated by an mRS score ranging from 0 to 1, while the remaining 2 patients (28.6%) showed incomplete functional recovery. The surgical interruption of CCJ-SDAVFs is the preferred treatment option, given its high obliteration rate and favorable functional recovery outcomes. We advocate the identification of C1 spinal nerve root as a crucial surgical step to identify the shunting points of CCJ- SDAVFs.

Global registration of kidneys in 3D ultrasound and CT images.

PURPOSE: Automatic registration between abdominal ultrasound (US) and computed tomography (CT) images is needed to enhance interventional guidance of renal procedures, but it remains an open research challenge. We propose a novel method that doesn't require an initial registration estimate (a global method) and also handles registration ambiguity caused by the organ's natural symmetry. Combined with a registration refinement algorithm, this method achieves robust and accurate kidney registration while avoiding manual initialization. METHODS: We propose solving global registration in a three-step approach: (1) Automatic anatomical landmark localization, where 2 deep neural networks (DNNs) localize a set of landmarks in each modality. (2) Registration hypothesis generation, where potential registrations are computed from the landmarks with a deterministic variant of RANSAC. Due to the Kidney's strong bilateral symmetry, there are usually 2 compatible solutions. Finally, in Step (3), the correct solution is determined automatically, using a DNN classifier that resolves the geometric ambiguity. The registration may then be iteratively improved with a registration refinement method. Results are presented with state-of-the-art surface-based refinement-Bayesian coherent point drift (BCPD). RESULTS: This automatic global registration approach gives better results than various competitive state-of-the-art methods, which, additionally, require organ segmentation. The results obtained on 59 pairs of 3D US/CT kidney images show that the proposed method, combined with BCPD refinement, achieves a target registration error (TRE) of an internal kidney landmark (the renal pelvis) of 5.78 mm and an average nearest neighbor surface distance (nndist) of 2.42 mm. CONCLUSION: This work presents the first approach for automatic kidney registration in US and CT images, which doesn't require an initial manual registration estimate to be known a priori. The results show a fully automatic registration approach with performances comparable to manual methods is feasible.

Comparison of two methods of greater occipital nerve block in patients with chronic migraine: ultrasound-guided and landmark-based techniques.

BACKGROUND: Migraine is a primary headache defined as moderate-to-severe pain lasting 4 to 72 h, ranking 2nd among the disabling conditions for both genders regardless of the age and the greater occipital nerve (GON) block has been reported as an efficient treatment method for migraine. The present study aims to evaluate and compare the efficiency of the two methods of GON block, i.e., the ultrasound (US)-guided technique and the landmark-based technique. METHOD: Having a prospective and randomized design, the study assigned the patients with chronic migraine into two groups after which a neurologist performed landmark-based GON block in the first group while an algologist performed US-guided GON block in the second group. During the 3-month follow-up period, the number of days with pain, the duration of pain, the number of analgesic drugs taken in a month, and Visual Analogue Scale (VAS) scores were compared with the values ​​before treatment and at the 1st week, 1st month, and 3rd month after treatment. RESULTS: US-guided GON block group included 34 patients while there were 32 patients in the landmark-based GON block group. US-guided GON block group showed significantly reduced VAS scores and frequency of attacks compared to the landmark-based GON block group at Month 1 after the procedure. After a 3-month follow-up period of the two groups, the frequency of attacks, analgesic intake and the duration of attacks were lower in both groups compared to the baseline. At 3-month follow-up, the mean of VAS scores decreased from 9,47 ± 2,69 to 4,67 ± 1,9 in US-guided GON block group and from 9,46 ± 0,98 to 7 ± 2,5 in the landmark-based GON block group. CONCLUSION: It was determined that both US-guided and landmark-based GON block were efficient techniques in patients with chronic migraine. US-guided GON block technique resulted in lower VAS scores, shorter durations of pain, lower frequencies of attack, and lower intake of analgesics compared to the landmark-based GON block technique.

Classification of cervical vertebral maturation stages with machine learning models: leveraging datasets with high inter- and intra-observer agreement.

OBJECTIVES: This study aimed to assess the accuracy of machine learning (ML) models with feature selection technique in classifying cervical vertebral maturation stages (CVMS). Consensus-based datasets were used for models training and evaluation for their model generalization capabilities on unseen datasets. METHODS: Three clinicians independently rated CVMS on 1380 lateral cephalograms, resulting in the creation of five datasets: two consensus-based datasets (Complete Agreement and Majority Voting), and three datasets based on a single rater's evaluations. Additionally, landmarks annotation of the second to fourth cervical vertebrae and patients' information underwent a feature selection process. These datasets were used to train various ML models and identify the top-performing model for each dataset. These models were subsequently tested on their generalization capabilities. RESULTS: Features that considered significant in the consensus-based datasets were consistent with a CVMS guideline. The Support Vector Machine model on the Complete Agreement dataset achieved the highest accuracy (77.4%), followed by the Multi-Layer Perceptron model on the Majority Voting dataset (69.6%). Models from individual ratings showed lower accuracies (60.4-67.9%). The consensus-based training models also exhibited lower coefficient of variation (CV), indicating superior generalization capability compared to models from single raters. CONCLUSION: ML models trained on consensus-based datasets for CVMS classification exhibited the highest accuracy, with significant features consistent with the original CVMS guidelines. These models also showed robust generalization capabilities, underscoring the importance of dataset quality.

Bayesian Landmark-based Shape Analysis of Tumor Pathology Images.

Medical imaging is a form of technology that has revolutionized the medical field over the past decades. Digital pathology imaging, which captures histological details at the cellular level, is rapidly becoming a routine clinical procedure for cancer diagnosis support and treatment planning. Recent developments in deep-learning methods have facilitated tumor region segmentation from pathology images. The traditional shape descriptors that characterize tumor boundary roughness at the anatomical level are no longer suitable. New statistical approaches to model tumor shapes are in urgent need. In this paper, we consider the problem of modeling a tumor boundary as a closed polygonal chain. A Bayesian landmark-based shape analysis model is proposed. The model partitions the polygonal chain into mutually exclusive segments, accounting for boundary roughness. Our Bayesian inference framework provides uncertainty estimations on both the number and locations of landmarks, while outputting metrics that can be used to quantify boundary roughness. The performance of our model is comparable with that of a recently developed landmark detection model for planar elastic curves. In a case study of 143 consecutive patients with stage I to IV lung cancer, we demonstrated the heterogeneity of tumor boundary roughness derived from our model effectively predicted patient prognosis (p-value < 0.001).

DetSegDiff: A joint periodontal landmark detection and segmentation in intraoral ultrasound using edge-enhanced diffusion-based network.

Individuals with malocclusion require an orthodontic diagnosis and treatment plan based on the severity of their condition. Assessing and monitoring changes in periodontal structures before, during, and after orthodontic procedures is crucial, and intraoral ultrasound (US) imaging has been shown a promising diagnostic tool in imaging periodontium. However, accurately delineating and analyzing periodontal structures in US videos is a challenging task for clinicians, as it is time-consuming and subject to interpretation errors. This paper introduces DetSegDiff, an edge-enhanced diffusion-based network developed to simultaneously detect the cementoenamel junction (CEJ) and segment alveolar bone structure in intraoral US videos. An edge feature encoder is designed to enhance edge and texture information for precise delineation of periodontal structures. Additionally, we employed the spatial squeeze-attention module (SSAM) to extract more representative features to perform both detection and segmentation tasks at global and local levels. This study used 169 videos from 17 orthodontic patients for training purposes and was subsequently tested on 41 videos from 4 additional patients. The proposed method achieved a mean distance difference of 0.17 ± 0.19 mm for the CEJ and an average Dice score of 90.1% for alveolar bone structure. As there is a lack of multi-task benchmark networks, thorough experiments were undertaken to assess and benchmark the proposed method against state-of-the-art (SOTA) detection and segmentation individual networks. The experimental results demonstrated that DetSegDiff outperformed SOTA approaches, confirming the feasibility of using automated diagnostic systems for orthodontists.

Implicit Is Not Enough: Explicitly Enforcing Anatomical Priors inside Landmark Localization Models.

The task of localizing distinct anatomical structures in medical image data is an essential prerequisite for several medical applications, such as treatment planning in orthodontics, bone-age estimation, or initialization of segmentation methods in automated image analysis tools. Currently, Anatomical Landmark Localization (ALL) is mainly solved by deep-learning methods, which cannot guarantee robust ALL predictions; there may always be outlier predictions that are far from their ground truth locations due to out-of-distribution inputs. However, these localization outliers are detrimental to the performance of subsequent medical applications that rely on ALL results. The current ALL literature relies heavily on implicit anatomical constraints built into the loss function and network architecture to reduce the risk of anatomically infeasible predictions. However, we argue that in medical imaging, where images are generally acquired in a controlled environment, we should use stronger explicit anatomical constraints to reduce the number of outliers as much as possible. Therefore, we propose the end-to-end trainable Global Anatomical Feasibility Filter and Analysis (GAFFA) method, which uses prior anatomical knowledge estimated from data to explicitly enforce anatomical constraints. GAFFA refines the initial localization results of a U-Net by approximately solving a Markov Random Field (MRF) with a single iteration of the sum-product algorithm in a differentiable manner. Our experiments demonstrate that GAFFA outperforms all other landmark refinement methods investigated in our framework. Moreover, we show that GAFFA is more robust to large outliers than state-of-the-art methods on the studied X-ray hand dataset. We further motivate this claim by visualizing the anatomical constraints used in GAFFA as spatial energy heatmaps, which allowed us to find an annotation error in the hand dataset not previously discussed in the literature.

Talar and Calcaneal Coordinate Axes Definitions across Foot Pathologies.

The understanding of foot and ankle biomechanics is improving as new technology provides more detailed information about the motion of foot and ankle bones with biplane fluoroscopy, as well as the ability to analyze the hindfoot under weightbearing conditions with weightbearing computed tomography. Three-dimensional anatomical coordinate systems are necessary to describe the 3D alignment and kinematics of the foot and ankle. The lack of standard coordinate systems across research study sites can significantly alter experimental data analyses used for pre-surgical evaluation and post-operative outcome assessments. Clinical treatment paradigms are changing based on the expanding knowledge of complex pes planovalgus morphologies or progressive collapsing foot deformity, which is present in both neurologic and non-neurologic populations. Four patient cohorts were created from 10 flexible PCFD, 10 rigid PCFD, 10 adult cerebral palsy, and 10 asymptomatic control patients. Six coordinate systems were tested on both the talus and calcaneus for all groups. The aim of this study was to evaluate axes definitions for the subtalar joint across four different patient populations to determine the influence of morphology on the implementation of previously defined coordinate systems. Different morphologic presentations from various pathologies have a substantial impact on coordinate system definitions, given that numerous axes definitions are defined through geometric fits or manual landmark selection. Automated coordinate systems that align with clinically relevant anatomic planes are preferred. Principal component axes are automatic, but do not align with clinically relevant planes and should not be used for such analysis where anatomic planes are critical.

AnyFace++: Deep Multi-Task, Multi-Domain Learning for Efficient Face AI.

Accurate face detection and subsequent localization of facial landmarks are mandatory steps in many computer vision applications, such as emotion recognition, age estimation, and gender identification. Thanks to advancements in deep learning, numerous facial applications have been developed for human faces. However, most have to employ multiple models to accomplish several tasks simultaneously. As a result, they require more memory usage and increased inference time. Also, less attention is paid to other domains, such as animals and cartoon characters. To address these challenges, we propose an input-agnostic face model, AnyFace++, to perform multiple face-related tasks concurrently. The tasks are face detection and prediction of facial landmarks for human, animal, and cartoon faces, including age estimation, gender classification, and emotion recognition for human faces. We trained the model using deep multi-task, multi-domain learning with a heterogeneous cost function. The experimental results demonstrate that AnyFace++ generates outcomes comparable to cutting-edge models designed for specific domains.

A Survey Study of the 3D Facial Landmark Detection Techniques Used as a Screening Tool for Diagnosis of the Obstructive Sleep Apnea Syndrome.

Obstructive Sleep Apnea (OSA) is a common disorder affecting both adults and children. It is characterized by repeated episodes of apnea (stopped breathing) and hypopnea (reduced breathing), which result in intermittent hypoxia. We recognize pediatric and adult OSA, and this paper focuses on pediatric OSA. While adults often suffer from daytime sleepiness, children are more likely to develop behavioral abnormalities. Early diagnosis and treatment are important to prevent negative effects on children's development. Without the treatment, children may be at increased risk of developing high blood pressure or other heart problems. The gold standard for OSA diagnosis is the polysomnography (sleep study) PSG performed at a sleep center. Not only is it an expensive procedure, but it can also be very stressful, especially for children. Patients have to stay at the sleep center during the night. Therefore, screening tools are very important. Multiple studies have shown that OSA screening tools can be based on facial anatomical landmarks. Anatomical landmarks are landmarks located at specific anatomical locations. For the purpose of the screening tool, a specific list of anatomical locations needs to be identified. We are presenting a survey study of the automatic identification of these landmarks on 3D scans of the patient's head. We are considering and comparing both knowledge-based and AI-based identification techniques, with a focus on the development of the automatic OSA screening tool.

Gait phase normalization resolves the problem of different phases being compared in gait cycle normalization.

For time-continuous analysis of gait, the problem of variations in cycle durations is resolved by normalizing to the gait cycle, but results depend on the definition of the cycle start. Gait cycle normalization ignores variations in gait phase durations, which results in averaging and comparing data across different phases. We propose gait phase normalization as part of a comprehensive method for independently analyzing magnitude and timing differences. First, gait phases are identified and differences in absolute and/or relative timing of phase durations or any point of interest between conditions or groups are analyzed using standard statistics. Next, time-continuous gait data is normalized to gait phases, and statistical parametric mapping (SPM) is used to assess magnitude differences in gait data. This approach is demonstrated on data recorded from ten young healthy adults walking on a treadmill at five different speeds. Sagittal knee angle was normalized to gait cycle or gait phase using five different gait cycle start events. Walking at different speeds resulted in significant changes in gait phase durations, highlighting a problem ignored by gait cycle normalization. SPM results for knee angle normalized to gait cycle varied from normalization to gait phases. Gait phase normalized SPM results were robust to the definition of the cycle start, in contrast to gait cycle normalized data. The approach of analyzing phase durations and normalizing data to gait phases overcomes previous limitations and enables a comprehensive analysis of magnitude and timing differences in time-continuous gait data and could be readily adapted to other tasks.

Heatmap-Based Active Shape Model for Landmark Detection in Lumbar X-ray Images.

Medical staff inspect lumbar X-ray images to diagnose lumbar spine diseases, and the analysis process is currently automated using deep-learning techniques. The detection of landmarks is necessary in the automatic process of localizing the position and identifying the morphological features of the vertebrae. However, detection errors may occur owing to the noise and ambiguity of images, as well as individual variations in the shape of the lumbar vertebrae. This study proposes a method to improve the robustness of landmark detection results. This method assumes that landmarks are detected by a convolutional neural network-based two-step model consisting of Pose-Net and M-Net. The model generates a heatmap response to indicate the probable landmark positions. The proposed method then corrects the landmark positions using the heatmap response and active shape model, which employs statistical information on the landmark distribution. Experiments were conducted using 3600 lumbar X-ray images, and the results showed that the landmark detection error was reduced by the proposed method. The average value of maximum errors decreased by 5.58% after applying the proposed method, which combines the outstanding image analysis capabilities of deep learning with statistical shape constraints on landmark distribution. The proposed method could also be easily integrated with other techniques to increase the robustness of landmark detection results such as CoordConv layers and non-directional part affinity field. This resulted in a further enhancement in the landmark detection performance. These advantages can improve the reliability of automatic systems used to inspect lumbar X-ray images. This will benefit both patients and medical staff by reducing medical expenses and increasing diagnostic efficiency.

Deep learning-based automated angle measurement for flatfoot diagnosis in weight-bearing lateral radiographs.

This study aimed to develop and evaluate a deep learning-based system for the automatic measurement of angles (specifically, Meary's angle and calcaneal pitch) in weight-bearing lateral radiographs of the foot for flatfoot diagnosis. We utilized 3960 lateral radiographs, either from the left or right foot, sourced from a pool of 4000 patients to construct and evaluate a deep learning-based model. These radiographs were captured between June and November 2021, and patients who had undergone total ankle replacement surgery or ankle arthrodesis surgery were excluded. Various methods, including correlation analysis, Bland-Altman plots, and paired T-tests, were employed to assess the concordance between the angles automatically measured using the system and those assessed by clinical experts. The evaluation dataset comprised 150 weight-bearing radiographs from 150 patients. In all test cases, the angles automatically computed using the deep learning-based system were in good agreement with the reference standards (Meary's angle: Pearson correlation coefficient (PCC) = 0.964, intraclass correlation coefficient (ICC) = 0.963, concordance correlation coefficient (CCC) = 0.963, p-value = 0.632, mean absolute error (MAE) = 1.59°; calcaneal pitch: PCC = 0.988, ICC = 0.987, CCC = 0.987, p-value = 0.055, MAE = 0.63°). The average time required for angle measurement using only the CPU to execute the deep learning-based system was 11 ± 1 s. The deep learning-based automatic angle measurement system, a tool for diagnosing flatfoot, demonstrated comparable accuracy and reliability with the results obtained by medical professionals for patients without internal fixation devices.

Artificial intelligence-assisted ultrasound-guided regional anaesthesia: An explorative scoping review.

Purpose: The present study reviews the available scientific literature on artificial intelligence (AI)-assisted ultrasound-guided regional anaesthesia (UGRA) and evaluates the reported intraprocedural parameters and postprocedural outcomes. Methods: A literature search was performed on 19 September 2023, using the Medline, EMBASE, CINAHL, Cochrane Library and Google Scholar databases by experts in electronic searching. All study designs were considered with no restrictions regarding patient characteristics or cohort size. Outcomes assessed included the accuracy of AI-model tracking, success at the first attempt, differences in outcomes between AI-assisted and unassisted UGRA, operator feedback and case-report data. Results: A joint adaptive median binary pattern (JAMBP) has been applied to improve the tracking procedure, while a particle filter (PF) is involved in feature extraction. JAMBP combined with PF was most accurate on all images for landmark identification, with accuracy scores of 0.83, 0.93 and 0.93 on original, preprocessed and filtered images, respectively. Evaluation of first-attempt success of spinal needle insertion revealed first-attempt success in most patients. When comparing AI application versus UGRA alone, a significant statistical difference (p < 0.05) was found for correct block view, correct structure identification and decrease in mean injection time, needle track adjustments and bone encounters in favour of having AI assistance. Assessment of operator feedback revealed that expert and nonexpert operator feedback was overall positive. Conclusion: AI appears promising to enhance UGRA as well as to positively influence operator training. AI application of UGRA may improve the identification of anatomical structures and provide guidance for needle placement, reducing the risk of complications and improving patient outcomes. Level of Evidence: Level IV.

Evaluation of automated photograph-cephalogram image integration using artificial intelligence models.

OBJECTIVES: To develop and evaluate an automated method for combining a digital photograph with a lateral cephalogram. MATERIALS AND METHODS: A total of 985 digital photographs were collected and soft tissue landmarks were manually detected. Then 2500 lateral cephalograms were collected, and corresponding soft tissue landmarks were manually detected. Using the images and landmark identification information, two different artificial intelligence (AI) models-one for detecting soft tissue on photographs and the other for identifying soft tissue on cephalograms-were developed using different deep-learning algorithms. The digital photographs were rotated, scaled, and shifted to minimize the squared sum of distances between the soft tissue landmarks identified by the two different AI models. As a validation process, eight soft tissue landmarks were selected on digital photographs and lateral cephalometric radiographs from 100 additionally collected validation subjects. Paired t-tests were used to compare the accuracy of measures obtained between the automated and manual image integration methods. RESULTS: The validation results showed statistically significant differences between the automated and manual methods on the upper lip and soft tissue B point. Otherwise, no statistically significant difference was found. CONCLUSIONS: Automated photograph-cephalogram image integration using AI models seemed to be as reliable as manual superimposition procedures.

Reinforcement learning-based anatomical maps for pancreas subregion and duct segmentation.

BACKGROUND: The pancreas is a complex abdominal organ with many anatomical variations, and therefore automated pancreas segmentation from medical images is a challenging application. PURPOSE: In this paper, we present a framework for segmenting individual pancreatic subregions and the pancreatic duct from three-dimensional (3D) computed tomography (CT) images. METHODS: A multiagent reinforcement learning (RL) network was used to detect landmarks of the head, neck, body, and tail of the pancreas, and landmarks along the pancreatic duct in a selected target CT image. Using the landmark detection results, an atlas of pancreases was nonrigidly registered to the target image, resulting in anatomical probability maps for the pancreatic subregions and duct. The probability maps were augmented with multilabel 3D U-Net architectures to obtain the final segmentation results. RESULTS: To evaluate the performance of our proposed framework, we computed the Dice similarity coefficient (DSC) between the predicted and ground truth manual segmentations on a database of 82 CT images with manually segmented pancreatic subregions and 37 CT images with manually segmented pancreatic ducts. For the four pancreatic subregions, the mean DSC improved from 0.38, 0.44, and 0.39 with standard 3D U-Net, Attention U-Net, and shifted windowing (Swin) U-Net architectures, to 0.51, 0.47, and 0.49, respectively, when utilizing the proposed RL-based framework. For the pancreatic duct, the RL-based framework achieved a mean DSC of 0.70, significantly outperforming the standard approaches and existing methods on different datasets. CONCLUSIONS: The resulting accuracy of the proposed RL-based segmentation framework demonstrates an improvement against segmentation with standard U-Net architectures.