Assessing the Accuracy of a Deep Learning Method to Risk Stratify Indeterminate Pulmonary Nodules.

Rationale: The management of indeterminate pulmonary nodules (IPNs) remains challenging, resulting in invasive procedures and delays in diagnosis and treatment. Strategies to decrease the rate of unnecessary invasive procedures and optimize surveillance regimens are needed.Objectives: To develop and validate a deep learning method to improve the management of IPNs.Methods: A Lung Cancer Prediction Convolutional Neural Network model was trained using computed tomography images of IPNs from the National Lung Screening Trial, internally validated, and externally tested on cohorts from two academic institutions.Measurements and Main Results: The areas under the receiver operating characteristic curve in the external validation cohorts were 83.5% (95% confidence interval [CI], 75.4-90.7%) and 91.9% (95% CI, 88.7-94.7%), compared with 78.1% (95% CI, 68.7-86.4%) and 81.9 (95% CI, 76.1-87.1%), respectively, for a commonly used clinical risk model for incidental nodules. Using 5% and 65% malignancy thresholds defining low- and high-risk categories, the overall net reclassifications in the validation cohorts for cancers and benign nodules compared with the Mayo model were 0.34 (Vanderbilt) and 0.30 (Oxford) as a rule-in test, and 0.33 (Vanderbilt) and 0.58 (Oxford) as a rule-out test. Compared with traditional risk prediction models, the Lung Cancer Prediction Convolutional Neural Network was associated with improved accuracy in predicting the likelihood of disease at each threshold of management and in our external validation cohorts.Conclusions: This study demonstrates that this deep learning algorithm can correctly reclassify IPNs into low- or high-risk categories in more than a third of cancers and benign nodules when compared with conventional risk models, potentially reducing the number of unnecessary invasive procedures and delays in diagnosis.

External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary nodules.

BACKGROUND: Estimation of the risk of malignancy in pulmonary nodules detected by CT is central in clinical management. The use of artificial intelligence (AI) offers an opportunity to improve risk prediction. Here we compare the performance of an AI algorithm, the lung cancer prediction convolutional neural network (LCP-CNN), with that of the Brock University model, recommended in UK guidelines. METHODS: A dataset of incidentally detected pulmonary nodules measuring 5-15 mm was collected retrospectively from three UK hospitals for use in a validation study. Ground truth diagnosis for each nodule was based on histology (required for any cancer), resolution, stability or (for pulmonary lymph nodes only) expert opinion. There were 1397 nodules in 1187 patients, of which 234 nodules in 229 (19.3%) patients were cancer. Model discrimination and performance statistics at predefined score thresholds were compared between the Brock model and the LCP-CNN. RESULTS: The area under the curve for LCP-CNN was 89.6% (95% CI 87.6 to 91.5), compared with 86.8% (95% CI 84.3 to 89.1) for the Brock model (p≤0.005). Using the LCP-CNN, we found that 24.5% of nodules scored below the lowest cancer nodule score, compared with 10.9% using the Brock score. Using the predefined thresholds, we found that the LCP-CNN gave one false negative (0.4% of cancers), whereas the Brock model gave six (2.5%), while specificity statistics were similar between the two models. CONCLUSION: The LCP-CNN score has better discrimination and allows a larger proportion of benign nodules to be identified without missing cancers than the Brock model. This has the potential to substantially reduce the proportion of surveillance CT scans required and thus save significant resources.

Correction to: Identifying Scoliosis in Population­Based Cohorts: Automation of a Validated Method Based on Total Body Dual Energy X­ray Absorptiometry Scans.

In the original version of the article, the co-author would like to add to the acknowledgements section to highlight their funding stream (EPSRC). The revised acknowledgements is given below.

Identifying Scoliosis in Population-Based Cohorts: Automation of a Validated Method Based on Total Body Dual Energy X-ray Absorptiometry Scans.

Scoliosis is a 3D-torsional rotation of the spine, but risk factors for initiation and progression are little understood. Research is hampered by lack of population-based research since radiographs cannot be performed on entire populations due to the relatively high levels of ionising radiation. Hence we have developed and validated a manual method for identifying scoliosis from total body dual energy X-ray absorptiometry (DXA) scans for research purposes. However, to allow full utilisation of population-based research cohorts, this needs to be automated. The purpose of this study was therefore to automate the identification of spinal curvature from total body DXA scans using machine learning techniques. To validate the automation, we assessed: (1) sensitivity, specificity and area under the receiver operator curve value (AUC) by comparison with 12,000 manually annotated images; (2) reliability by rerunning the automation on a subset of DXA scans repeated 2-6 weeks apart and calculating the kappa statistic; (3) validity by applying the automation to 5000 non-annotated images to assess associations with epidemiological variables. The final automated model had a sensitivity of 86.5%, specificity of 96.9% and an AUC of 0.80 (95%CI 0.74-0.87). There was almost perfect agreement of identification of those with scoliosis (kappa 0.90). Those with scoliosis identified by the automated model showed similar associations with gender, ethnicity, socioeconomic status, BMI and lean mass to previous literature. In conclusion, we have developed an accurate and valid automated method for identifying and quantifying spinal curvature from total body DXA scans.

Could automated machine-learned MRI grading aid epidemiological studies of lumbar spinal stenosis? Validation within the Wakayama spine study.

BACKGROUND: MRI scanning has revolutionized the clinical diagnosis of lumbar spinal stenosis (LSS). However, there is currently no consensus as to how best to classify MRI findings which has hampered the development of robust longitudinal epidemiological studies of the condition. We developed and tested an automated system for grading lumbar spine MRI scans for central LSS for use in epidemiological research. METHODS: Using MRI scans from the large population-based cohort study (the Wakayama Spine Study), all graded by a spinal surgeon, we trained an automated system to grade central LSS in four gradings of the bone and soft tissue margins: none, mild, moderate, severe. Subsequently, we tested the automated grading against the independent readings of our observer in a test set to investigate reliability and agreement. RESULTS: Complete axial views were available for 4855 lumbar intervertebral levels from 971 participants. The machine used 4365 axial views to learn (training set) and graded the remaining 490 axial views (testing set). The agreement rate for gradings was 65.7% (322/490) and the reliability (Lin's correlation coefficient) was 0.73. In 2.2% of scans (11/490) there was a difference in classification of 2 and in only 0.2% (1/490) was there a difference of 3. When classified into 2 groups as 'severe' vs 'no/mild/moderate'. The agreement rate was 94.1% (461/490) with a kappa of 0.75. CONCLUSIONS: This study showed that an automated system can "learn" to grade central LSS with excellent performance against the reference standard. Thus SpineNet offers potential to grade LSS in large-scale epidemiological studies involving a high volume of MRI spine data with a high level of consistency and objectivity.

ISSLS PRIZE IN BIOENGINEERING SCIENCE 2017: Automation of reading of radiological features from magnetic resonance images (MRIs) of the lumbar spine without human intervention is comparable with an expert radiologist.

STUDY DESIGN: Investigation of the automation of radiological features from magnetic resonance images (MRIs) of the lumbar spine. OBJECTIVE: To automate the process of grading lumbar intervertebral discs and vertebral bodies from MRIs. MR imaging is the most common imaging technique used in investigating low back pain (LBP). Various features of degradation, based on MRIs, are commonly recorded and graded, e.g., Modic change and Pfirrmann grading of intervertebral discs. Consistent scoring and grading is important for developing robust clinical systems and research. Automation facilitates this consistency and reduces the time of radiological analysis considerably and hence the expense. METHODS: 12,018 intervertebral discs, from 2009 patients, were graded by a radiologist and were then used to train: (1) a system to detect and label vertebrae and discs in a given scan, and (2) a convolutional neural network (CNN) model that predicts several radiological gradings. The performance of the model, in terms of class average accuracy, was compared with the intra-observer class average accuracy of the radiologist. RESULTS: The detection system achieved 95.6% accuracy in terms of disc detection and labeling. The model is able to produce predictions of multiple pathological gradings that consistently matched those of the radiologist. The model identifies 'Evidence Hotspots' that are the voxels that most contribute to the degradation scores. CONCLUSIONS: Automation of radiological grading is now on par with human performance. The system can be beneficial in aiding clinical diagnoses in terms of objectivity of gradings and the speed of analysis. It can also draw the attention of a radiologist to regions of degradation. This objectivity and speed is an important stepping stone in the investigation of the relationship between MRIs and clinical diagnoses of back pain in large cohorts. LEVEL OF EVIDENCE: Level 3.

Automated detection, labelling and radiological grading of clinical spinal MRIs.

Spinal magnetic resonance (MR) scans are a vital tool for diagnosing the cause of back pain for many diseases and conditions. However, interpreting clinically useful information from these scans can be challenging, time-consuming and hard to reproduce across different radiologists. In this paper, we alleviate these problems by introducing a multi-stage automated pipeline for analysing spinal MR scans. This pipeline first detects and labels vertebral bodies across several commonly used sequences (e.g. T1w, T2w and STIR) and fields of view (e.g. lumbar, cervical, whole spine). Using these detections it then performs automated diagnosis for several spinal disorders, including intervertebral disc degenerative changes in T1w and T2w lumbar scans, and spinal metastases, cord compression and vertebral fractures. To achieve this, we propose a new method of vertebrae detection and labelling, using vector fields to group together detected vertebral landmarks and a language-modelling inspired beam search to determine the corresponding levels of the detections. We also employ a new transformer-based architecture to perform radiological grading which incorporates context from multiple vertebrae and sequences, as a real radiologist would. The performance of each stage of the pipeline is tested in isolation on several clinical datasets, each consisting of 66 to 421 scans. The outputs are compared to manual annotations of expert radiologists, demonstrating accurate vertebrae detection across a range of scan parameters. Similarly, the model's grading predictions for various types of disc degeneration and detection of spinal metastases closely match those of an expert radiologist. To aid future research, our code and trained models are made publicly available.

Automated measurement of size of spinal curve in population-based cohorts: Validation of a method based on total body dual energy X-ray absorptiometry scans.

BACKGROUND: Scoliosis is spinal curvature that may progress to require surgical stabilisation. Risk factors for progression are little understood due to lack of population-based research, since radiographs cannot be performed on entire populations due to high levels of radiation. To help address this, we have previously developed and validated a method for quantification of spinal curvature from total body dual energy X-ray absorptiometry (DXA) scans. The purpose of this study was to automate this quantification of spinal curve size from DXA scans using machine learning techniques. METHODS: To develop the automation of curve size, we utilised manually annotated scans from 7298 participants from the Avon Longitudinal Study of Parents and Children (ALSPAC) at age 9 and 5122 at age 15. To validate the automation we assessed (1) agreement between manual vs automation using the Bland-Altman limits of agreement, (2) reliability by calculating the coefficient of variation, and (3) clinical validity by running the automation on 4969 non-annotated scans at age 18 to assess the associations with physical activity, body composition, adipocyte function and backpain compared to previous literature. RESULTS: The mean difference between manual vs automated readings was less than one degree, and 90.4 % of manual vs automated readings fell within 10°. The coefficient of variation was 25.4 %. Clinical validation showed the expected relationships between curve size and physical activity, adipocyte function, height and weight. CONCLUSION: We have developed a reasonably accurate and valid automated method for quantifying spinal curvature from DXA scans for research purposes.

External Validation of SpineNet, an Open-Source Deep Learning Model for Grading Lumbar Disk Degeneration MRI Features, Using the Northern Finland Birth Cohort 1966.

STUDY DESIGN: This is a retrospective observational study to externally validate a deep learning image classification model. OBJECTIVE: Deep learning models such as SpineNet offer the possibility of automating the process of disk degeneration (DD) classification from magnetic resonance imaging (MRI). External validation is an essential step to their development. The aim of this study was to externally validate SpineNet predictions for DD using Pfirrmann classification and Modic changes (MCs) on data from the Northern Finland Birth Cohort 1966 (NFBC1966). SUMMARY OF DATA: We validated SpineNet using data from 1331 NFBC1966 participants for whom both lumbar spine MRI data and consensus DD gradings were available. MATERIALS AND METHODS: SpineNet returned Pfirrmann grade and MC presence from T2-weighted sagittal lumbar MRI sequences from NFBC1966, a data set geographically and temporally separated from its training data set. A range of agreement and reliability metrics were used to compare predictions with expert radiologists. Subsets of data that match SpineNet training data more closely were also tested. RESULTS: Balanced accuracy for DD was 78% (77%-79%) and for MC 86% (85%-86%). Interrater reliability for Pfirrmann grading was Lin concordance correlation coefficient=0.86 (0.85-0.87) and Cohen κ=0.68 (0.67-0.69). In a low back pain subset, these reliability metrics remained largely unchanged. In total, 20.83% of disks were rated differently by SpineNet compared with the human raters, but only 0.85% of disks had a grade difference >1. Interrater reliability for MC detection was κ=0.74 (0.72-0.75). In the low back pain subset, this metric was almost unchanged at κ=0.76 (0.73-0.79). CONCLUSIONS: In this study, SpineNet has been benchmarked against expert human raters in the research setting. It has matched human reliability and demonstrates robust performance despite the multiple challenges facing model generalizability.

A deep learning approach to private data sharing of medical images using conditional generative adversarial networks (GANs).

Clinical data sharing can facilitate data-driven scientific research, allowing a broader range of questions to be addressed and thereby leading to greater understanding and innovation. However, sharing biomedical data can put sensitive personal information at risk. This is usually addressed by data anonymization, which is a slow and expensive process. An alternative to anonymization is construction of a synthetic dataset that behaves similar to the real clinical data but preserves patient privacy. As part of a collaboration between Novartis and the Oxford Big Data Institute, a synthetic dataset was generated based on images from COSENTYX® (secukinumab) ankylosing spondylitis (AS) clinical studies. An auxiliary classifier Generative Adversarial Network (ac-GAN) was trained to generate synthetic magnetic resonance images (MRIs) of vertebral units (VUs), conditioned on the VU location (cervical, thoracic and lumbar). Here, we present a method for generating a synthetic dataset and conduct an in-depth analysis on its properties along three key metrics: image fidelity, sample diversity and dataset privacy.

The utility of a convolutional neural network (CNN) model score for cancer risk in indeterminate small solid pulmonary nodules, compared to clinical practice according to British Thoracic Society guidelines.

PURPOSE: To determine how implementation of an artificial intelligence nodule algorithm, the Lung Cancer Prediction Convolutional Neural Network (LCP-CNN), at the point of incidental nodule detection would have influenced further investigation and management using a series of threshold scores at both the benign and malignant end of the spectrum. METHOD: An observational retrospective study was performed in the assessment of nodules between 5-15 mm (158 benign, 32 malignant) detected on CT scans, which were performed as part of routine practice. The LCP-CNN was applied to the baseline CT scan producing a percentage score, and subsequent imaging and management determined for each threshold group. We hypothesized that the 5% low risk threshold group requires only one follow-up, the 0.56% very low risk threshold group requires no follow-up and the 80% high risk threshold group warrants expedited intervention. RESULTS: The 158 benign nodules had an LCP-CNN score between 0.1 and 70.8%, median 5.5% (IQR 1.4-18.0), whilst the 32 cancer nodules had an LCP-CNN score between 10.1 and 98.7%, median 59.0% (IQR 37.1-83.9). 24/61 CT scans in the 0.56-5% group (n = 37) and 21/21 CT scans <0.56% group (n = 13) could be obviated resulting in an overall reduction of 18.6% (45/242) CT scans in the benign cohort. In the 80% group (n = 10), expedited intervention of malignant nodules could result in a 3.6-month reduction in time delay in 5 cancer patients. CONCLUSION: We show the potential of artificial intelligence to reduce the need for follow-up scans and intervention in low-scoring benign nodules, whilst potentially accelerating the investigation and treatment of high-scoring cancer nodules.

Lung cancer prediction by Deep Learning to identify benign lung nodules.

INTRODUCTION: Deep Learning has been proposed as promising tool to classify malignant nodules. Our aim was to retrospectively validate our Lung Cancer Prediction Convolutional Neural Network (LCP-CNN), which was trained on US screening data, on an independent dataset of indeterminate nodules in an European multicentre trial, to rule out benign nodules maintaining a high lung cancer sensitivity. METHODS: The LCP-CNN has been trained to generate a malignancy score for each nodule using CT data from the U.S. National Lung Screening Trial (NLST), and validated on CT scans containing 2106 nodules (205 lung cancers) detected in patients from from the Early Lung Cancer Diagnosis Using Artificial Intelligence and Big Data (LUCINDA) study, recruited from three tertiary referral centers in the UK, Germany and Netherlands. We pre-defined a benign nodule rule-out test, to identify benign nodules whilst maintaining a high sensitivity, by calculating thresholds on the malignancy score that achieve at least 99 % sensitivity on the NLST data. Overall performance per validation site was evaluated using Area-Under-the-ROC-Curve analysis (AUC). RESULTS: The overall AUC across the European centers was 94.5 % (95 %CI 92.6-96.1). With a high sensitivity of 99.0 %, malignancy could be ruled out in 22.1 % of the nodules, enabling 18.5 % of the patients to avoid follow-up scans. The two false-negative results both represented small typical carcinoids. CONCLUSION: The LCP-CNN, trained on participants with lung nodules from the US NLST dataset, showed excellent performance on identification of benign lung nodules in a multi-center external dataset, ruling out malignancy with high accuracy in about one fifth of the patients with 5-15 mm nodules.

The use of technology in the subcategorisation of osteoarthritis: a Delphi study approach.

Objective: This UK-wide OATech Network + consensus study utilised a Delphi approach to discern levels of awareness across an expert panel regarding the role of existing and novel technologies in osteoarthritis research. To direct future cross-disciplinary research it aimed to identify which could be adopted to subcategorise patients with osteoarthritis (OA). Design: An online questionnaire was formulated based on technologies which might aid OA research and subcategorisation. During a two-day face-to-face meeting concordance of expert opinion was established with surveys (23 questions) before, during and at the end of the meeting (Rounds 1, 2 and 3, respectively). Experts spoke on current evidence for imaging, genomics, epigenomics, proteomics, metabolomics, biomarkers, activity monitoring, clinical engineering and machine learning relating to subcategorisation. For each round of voting, ≥80% votes led to consensus and ≤20% to exclusion of a statement. Results: Panel members were unanimous that a combination of novel technological advances have potential to improve OA diagnostics and treatment through subcategorisation, agreeing in Rounds 1 and 2 that epigenetics, genetics, MRI, proteomics, wet biomarkers and machine learning could aid subcategorisation. Expert presentations changed participants' opinions on the value of metabolomics, activity monitoring and clinical engineering, all reaching consensus in Round 2. X-rays lost consensus between Rounds 1 and 2; clinical X-rays reached consensus in Round 3. Conclusion: Consensus identified that 9 of the 11 technologies should be targeted towards OA subcategorisation to address existing OA research technology and knowledge gaps. These novel, rapidly evolving technologies are recommended as a focus for emergent, cross-disciplinary osteoarthritis research programmes.

Can Atlas-Based Auto-Segmentation Ever Be Perfect? Insights From Extreme Value Theory.

Atlas-based segmentation is used in radiotherapy planning to accelerate the delineation of organs at risk (OARs). Atlas selection has been proposed to improve the performance of segmentation, assuming that the more similar the atlas is to the patient, the better the result. It follows that the larger the database of atlases from which to select, the better the results should be. This paper seeks to estimate a clinically achievable expected performance under this assumption. Assuming a perfect atlas selection, an extreme value theory has been applied to estimate the accuracy of single-atlas and multi-atlas segmentation given a large database of atlases. For this purpose, clinical contours of most common OARs on computed tomography of the head and neck ( N=316 ) and thoracic ( N=280 ) cases were used. This paper found that while for most organs, perfect segmentation cannot be reasonably expected, auto-contouring performance of a level corresponding to clinical quality could be consistently expected given a database of 5000 atlases under the assumption of perfect atlas selection.

Estimating the joint statistics of images using nonparametric windows with application to registration using mutual information.

Recently, the Nonparametric (NP) Windows has been proposed to estimate the statistics of real 1D and 2D signals. NP Windows is accurate, because it is equivalent to sampling images at a high (infinite) resolution for an assumed interpolation model. This paper extends the proposed approach to consider joint distributions of image-pairs. Second, Green's Theorem is used to simplify the previous NP Windows algorithm. Finally, a resolution-aware NP Windows algorithm is proposed to improve robustness to relative scaling between an image pair. Comparative testing of 2D image registration was performed using translation-only and affine transformations. Although more expensive than other methods, NP Windows frequently demonstrated superior performance for bias (distance between ground truth and global maximum) and frequency of convergence. Unlike other methods, the number of samples and the number of bins have little effect on NP Windows and the prior selection of a kernel is not required.

Lung cancer prediction using machine learning and advanced imaging techniques.

Machine learning based lung cancer prediction models have been proposed to assist clinicians in managing incidental or screen detected indeterminate pulmonary nodules. Such systems may be able to reduce variability in nodule classification, improve decision making and ultimately reduce the number of benign nodules that are needlessly followed or worked-up. In this article, we provide an overview of the main lung cancer prediction approaches proposed to date and highlight some of their relative strengths and weaknesses. We discuss some of the challenges in the development and validation of such techniques and outline the path to clinical adoption.

SpineNet: Automated classification and evidence visualization in spinal MRIs.

The objective of this work is to automatically produce radiological gradings of spinal lumbar MRIs and also localize the predicted pathologies. We show that this can be achieved via a Convolutional Neural Network (CNN) framework that takes intervertebral disc volumes as inputs and is trained only on disc-specific class labels. Our contributions are: (i) a CNN architecture that predicts multiple gradings at once, and we propose variants of the architecture including using 3D convolutions; (ii) showing that this architecture can be trained using a multi-task loss function without requiring segmentation level annotation; and (iii) a localization method that clearly shows pathological regions in the disc volumes. We compare three visualization methods for the localization. The network is applied to a large corpus of MRI T2 sagittal spinal MRIs (using a standard clinical scan protocol) acquired from multiple machines, and is used to automatically compute disk and vertebra gradings for each MRI. These are: Pfirrmann grading, disc narrowing, upper/lower endplate defects, upper/lower marrow changes, spondylolisthesis, and central canal stenosis. We report near human performances across the eight gradings, and also visualize the evidence for these gradings localized on the original scans.

Validation of clinical acceptability of an atlas-based segmentation algorithm for the delineation of organs at risk in head and neck cancer.

PURPOSE: The aim of this study was to assess whether clinically acceptable segmentations of organs at risk (OARs) in head and neck cancer can be obtained automatically and efficiently using the novel "similarity and truth estimation for propagated segmentations" (STEPS) compared to the traditional "simultaneous truth and performance level estimation" (STAPLE) algorithm. METHODS: First, 6 OARs were contoured by 2 radiation oncologists in a dataset of 100 patients with head and neck cancer on planning computed tomography images. Each image in the dataset was then automatically segmented with STAPLE and STEPS using those manual contours. Dice similarity coefficient (DSC) was then used to compare the accuracy of these automatic methods. Second, in a blind experiment, three separate and distinct trained physicians graded manual and automatic segmentations into one of the following three grades: clinically acceptable as determined by universal delineation guidelines (grade A), reasonably acceptable for clinical practice upon manual editing (grade B), and not acceptable (grade C). Finally, STEPS segmentations graded B were selected and one of the physicians manually edited them to grade A. Editing time was recorded. RESULTS: Significant improvements in DSC can be seen when using the STEPS algorithm on large structures such as the brainstem, spinal canal, and left/right parotid compared to the STAPLE algorithm (all p < 0.001). In addition, across all three trained physicians, manual and STEPS segmentation grades were not significantly different for the brainstem, spinal canal, parotid (right/left), and optic chiasm (all p > 0.100). In contrast, STEPS segmentation grades were lower for the eyes (p < 0.001). Across all OARs and all physicians, STEPS produced segmentations graded as well as manual contouring at a rate of 83%, giving a lower bound on this rate of 80% with 95% confidence. Reduction in manual interaction time was on average 61% and 93% when automatic segmentations did and did not, respectively, require manual editing. CONCLUSIONS: The STEPS algorithm showed better performance than the STAPLE algorithm in segmenting OARs for radiotherapy of the head and neck. It can automatically produce clinically acceptable segmentation of OARs, with results as relevant as manual contouring for the brainstem, spinal canal, the parotids (left/right), and optic chiasm. A substantial reduction in manual labor was achieved when using STEPS even when manual editing was necessary.

Learning new parts for landmark localization in whole-body CT scans.

The goal of this work is to reliably and accurately localize anatomical landmarks in 3-D computed tomography scans, particularly for the deformable registration of whole-body scans, which show huge variation in posture, and the spatial distribution of anatomical features. Parts-based graphical models (GM) have shown attractive properties for this task because they capture naturally anatomical relationships between landmarks. Unfortunately, standard GMs are learned from manually annotated training images and the quantity of landmarks is limited by the high cost of expert annotation. We propose a novel method that automatically learns new corresponding landmarks from a database of 3-D whole-body CT scans, using a limited initial set of expert-labeled ground-truth landmarks. The newly learned landmarks, called B-landmarks, are used to build enriched GMs. We compare our method of deformable registration based on such GM landmarks to a conventional deformable registration method and to a "baseline" state-of-the-art GM. The results show our method finds new relevant anatomical correspondences and improves by up to 35% the matching accuracy of highly variable skeletal and soft-tissue landmarks of clinical interest.