Learning 3D medical image keypoint descriptors with the triplet loss.

PURPOSE: We propose to learn a 3D keypoint descriptor which we use to match keypoints extracted from full-body CT scans. Our methods are inspired by 2D keypoint descriptor learning, which was shown to outperform hand-crafted descriptors. Adapting these to 3D images is challenging because of the lack of labelled training data and high memory requirements. METHOD: We generate semi-synthetic training data. For that, we first estimate the distribution of local affine inter-subject transformations using labelled anatomical landmarks on a small subset of the database. We then sample a large number of transformations and warp unlabelled CT scans, for which we can subsequently establish reliable keypoint correspondences using guided matching. These correspondences serve as training data for our descriptor, which we represent by a CNN and train using the triplet loss with online triplet mining. RESULTS: We carry out experiments on a synthetic data reliability benchmark and a registration task involving 20 CT volumes with anatomical landmarks used for evaluation purposes. Our learned descriptor outperforms the 3D-SURF descriptor on both benchmarks while having a similar runtime. CONCLUSION: We propose a new method to generate semi-synthetic data and a new learned 3D keypoint descriptor. Experiments show improvement compared to a hand-crafted descriptor. This is promising as literature has shown that jointly learning a detector and a descriptor gives further performance boost.

Automatic Multiorgan Segmentation via Multiscale Registration and Graph Cut.

We propose an automatic multiorgan segmentation method for 3-D radiological images of different anatomical contents and modalities. The approach is based on a simultaneous multilabel graph cut optimization of location, appearance, and spatial configuration criteria of target structures. Organ location is defined by target-specific probabilistic atlases (PA) constructed from a training dataset using a fast (2+1)D SURF-based multiscale registration method involving a simple four-parameter transformation. PAs are also used to derive target-specific organ appearance models represented as intensity histograms. The spatial configuration prior is derived from shortest-path constraints defined on the adjacency graph of structures. Thorough evaluations on Visceral project benchmarks and training dataset, as well as comparisons with the state-of-the-art confirm that our approach is comparable to and often outperforms similar approaches in multiorgan segmentation, thus proving that the combination of multiple suboptimal but complementary information sources can yield very good performance.

Cloud-Based Evaluation of Anatomical Structure Segmentation and Landmark Detection Algorithms: VISCERAL Anatomy Benchmarks.

Variations in the shape and appearance of anatomical structures in medical images are often relevant radiological signs of disease. Automatic tools can help automate parts of this manual process. A cloud-based evaluation framework is presented in this paper including results of benchmarking current state-of-the-art medical imaging algorithms for anatomical structure segmentation and landmark detection: the VISCERAL Anatomy benchmarks. The algorithms are implemented in virtual machines in the cloud where participants can only access the training data and can be run privately by the benchmark administrators to objectively compare their performance in an unseen common test set. Overall, 120 computed tomography and magnetic resonance patient volumes were manually annotated to create a standard Gold Corpus containing a total of 1295 structures and 1760 landmarks. Ten participants contributed with automatic algorithms for the organ segmentation task, and three for the landmark localization task. Different algorithms obtained the best scores in the four available imaging modalities and for subsets of anatomical structures. The annotation framework, resulting data set, evaluation setup, results and performance analysis from the three VISCERAL Anatomy benchmarks are presented in this article. Both the VISCERAL data set and Silver Corpus generated with the fusion of the participant algorithms on a larger set of non-manually-annotated medical images are available to the research community.

Shortest-path constraints for 3D multiobject semiautomatic segmentation via clustering and Graph Cut.

We derive shortest-path constraints from graph models of structure adjacency relations and introduce them in a joint centroidal Voronoi image clustering and Graph Cut multiobject semiautomatic segmentation framework. The vicinity prior model thus defined is a piecewise-constant model incurring multiple levels of penalization capturing the spatial configuration of structures in multiobject segmentation. Qualitative and quantitative analyses and comparison with a Potts prior-based approach and our previous contribution on synthetic, simulated, and real medical images show that the vicinity prior allows for the correct segmentation of distinct structures having identical intensity profiles and improves the precision of segmentation boundary placement while being fairly robust to clustering resolution. The clustering approach we take to simplify images prior to segmentation strikes a good balance between boundary adaptivity and cluster compactness criteria furthermore allowing to control the trade-off. Compared with a direct application of segmentation on voxels, the clustering step improves the overall runtime and memory footprint of the segmentation process up to an order of magnitude without compromising the quality of the result.