Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Curve Detection in Volumetric Medical Imaging via Attraction Field

Farukh Yaushev, Daria Nogina, Valentin Samokhin, Mariya Dugova, Ekaterina Petrash, Dmitry Sevryukov, Mikhail Belyaev, Maxim Pisov

TL;DR

This work tackles robust detection of non-branching curves in volumetric medical images, a task where segmentation-based methods struggle to capture subpixel-accurate curves such as the aortic centerline or vertebral column centerline. It introduces a 3D two-headed CNN that jointly predicts an attraction field guiding voxels to the curve and a closeness map to constrain the ROI, with a loss that combines $L_{field}$, $L_{cls}$, and $L_{norm}$ to stabilize training. Inference converts the resulting voxel-wise predictions into a thinned point cloud, which is then ordered into a 1D curve using non-maximum suppression and Isomap for parameterization, achieving subpixel accuracy. The method shows superior performance on both aortic and vertebral centerline tasks across multiple metrics and provides public release of annotations as a benchmark, promoting reproducibility and future research. Although focused on non-branching curves, the authors suggest extensions to branching structures via branching-point prediction, highlighting practical impact for medical imaging diagnostics.

Abstract

Understanding body part geometry is crucial for precise medical diagnostics. Curves effectively describe anatomical structures and are widely used in medical imaging applications related to cardiovascular, respiratory, and skeletal diseases. Traditional curve detection methods are often task-specific, relying heavily on domain-specific features, limiting their broader applicability. This paper introduces a novel approach for detecting non-branching curves, which does not require prior knowledge of the object's orientation, shape, or position. Our method uses neural networks to predict (1) an attraction field, which offers subpixel accuracy, and (2) a closeness map, which limits the region of interest and essentially eliminates outliers far from the desired curve. We tested our curve detector on several clinically relevant tasks with diverse morphologies and achieved impressive subpixel-level accuracy results that surpass existing methods, highlighting its versatility and robustness. Additionally, to support further advancements in this field, we provide our private annotations of aortic centerlines and masks, which can serve as a benchmark for future research. The dataset can be found at https://github.com/neuro-ml/curve-detection.

Robust Curve Detection in Volumetric Medical Imaging via Attraction Field

TL;DR

This work tackles robust detection of non-branching curves in volumetric medical images, a task where segmentation-based methods struggle to capture subpixel-accurate curves such as the aortic centerline or vertebral column centerline. It introduces a 3D two-headed CNN that jointly predicts an attraction field guiding voxels to the curve and a closeness map to constrain the ROI, with a loss that combines , , and to stabilize training. Inference converts the resulting voxel-wise predictions into a thinned point cloud, which is then ordered into a 1D curve using non-maximum suppression and Isomap for parameterization, achieving subpixel accuracy. The method shows superior performance on both aortic and vertebral centerline tasks across multiple metrics and provides public release of annotations as a benchmark, promoting reproducibility and future research. Although focused on non-branching curves, the authors suggest extensions to branching structures via branching-point prediction, highlighting practical impact for medical imaging diagnostics.

Abstract

Understanding body part geometry is crucial for precise medical diagnostics. Curves effectively describe anatomical structures and are widely used in medical imaging applications related to cardiovascular, respiratory, and skeletal diseases. Traditional curve detection methods are often task-specific, relying heavily on domain-specific features, limiting their broader applicability. This paper introduces a novel approach for detecting non-branching curves, which does not require prior knowledge of the object's orientation, shape, or position. Our method uses neural networks to predict (1) an attraction field, which offers subpixel accuracy, and (2) a closeness map, which limits the region of interest and essentially eliminates outliers far from the desired curve. We tested our curve detector on several clinically relevant tasks with diverse morphologies and achieved impressive subpixel-level accuracy results that surpass existing methods, highlighting its versatility and robustness. Additionally, to support further advancements in this field, we provide our private annotations of aortic centerlines and masks, which can serve as a benchmark for future research. The dataset can be found at https://github.com/neuro-ml/curve-detection.
Paper Structure (20 sections, 5 equations, 4 figures, 2 tables)

This paper contains 20 sections, 5 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Loss Function
  4. Attraction Field
  5. Closeness Map
  6. Handling multiple projections
  7. Final Loss
  8. Inference
  9. Experimental Setup
  10. Data
  11. Training Details
  12. Metrics
  13. Baselines
  14. Results
  15. Models Performance
  16. ...and 5 more sections

Figures (4)

  • Figure 1: A schematic representation of the prediction pipeline for aortic centerline detection: a) a $3$D input image, red circles denote the aorta; b) the predicted attraction field depicted for a given $2$D slice, the red color indicates the field vectors that lead to inaccurate predictions; c) the closeness map predicted for a given $2$D slice; d) the same $2$D slice relative to the predicted point cloud, the black dots represent the unordered set of points; e) the result of ordering using Isomap -- the final predicted $3$D curve.
  • Figure 2: Aortic centerline predictions for various methods in sagittal or coronal projections, with the aorta highlighted in each case. Only relevant curves are displayed for clarity. a) Typical prediction without major defects; b) Magnified region showing the roughness of the Seg model's predicted curve; c) Example of a false-positive for the Att model caused by another tubular structure; d) Erroneous prediction by our method; e) Axial slice where our method did not make a prediction due to noisiness - the aorta is indistinguishable.
  • Figure 3: Vertebral column centerline predictions for several methods in sagittal projections. For clarity, only curves of interest are shown. a) a typical prediction without major errors from our method; b) a magnified region highlighting the roughness of the predicted curve by the Seg model; c) a magnified region highlighting the roughness of the predicted curve by the Htmp model; d) an example of a typical false-positive for the Att model caused by another tubular structure; e) an example of an erroneous prediction by our method.
  • Figure 4: A sagittal projection of an image with kyphosis. The red region indicates the range of axial slices intersecting the curve at two locations -- the cause of the inapplicability of soft-argmax-based methods. The green curve represents the prediction generated by our model.