Disentangling Overlapping Anatomical Structures in Radiographic Imaging

47 concepts42 papers2 proposals

When a three-dimensional human body is projected onto a two-dimensional radiographic plane, depth information is lost. High-density skeletal structures, soft tissues, and blood vessels are compressed into a single layer of pixels. This phenomenon, known as anatomical occlusion, creates a significant challenge for both human practitioners and computer vision models.

A lung nodule, a thin pleural line indicating a collapsed lung, or a misplaced medical device can easily be masked by the shadow of a rib or a clavicle. To build automated diagnostic systems that rival expert radiologists, researchers have developed techniques to computationally "un-layer" these projections, isolating pathologies from the healthy structures that obscure them.


The Geometry of Projection and Why It Breaks Vision Models

Standard convolutional neural networks (CNNs) rely heavily on local texture gradients, contrast boundaries, and edge detection to segment organs or identify anomalies. Anatomical occlusion disrupts these visual cues in two distinct ways:

  1. Shadow Mimicry: Dense structures produce shadows that closely resemble pathological opacities. As noted early on by Harry Wechsler et al. in Automatic Detection Of Rib Contours in Chest Radiographs, skeletal contours can mimic the appearance of tumors, leading to high false-positive rates in automated detection pipelines. Similarly, F. Pasa et al. demonstrated in Efficient Deep Network Architectures for Fast Chest X-Ray Tuberculosis Screening and Visualization that bone superposition artifacts frequently confuse screening models by mimicking tuberculosis manifestations.
  2. Boundary Leakage: When a high-contrast bone overlaps a low-contrast soft-tissue boundary, the model's edge-detection filters track the sharper skeletal edge. Lingdong Chen et al. showed in Development of lung segmentation method in x-ray images of children based on TransResUNet that this lack of clear contrast causes segmentation boundaries to "leak," resulting in jagged, inaccurate organ margins.

This problem is not unique to chest X-rays. Whether analyzing the millimeter-scale internal components of the temporal bone (Yoshiaki Nakai et al. in Tomography of the Temporal Bone), locating landmarks on the lower spine obscured by pelvic bones (Shuo Yuan et al. in Artificial intelligence automatic measurement technology of lumbosacral radiographic parameters), or tracking blood vessels behind dense structures (Alina F. Dima et al. in 3D Arterial Segmentation via Single 2D Projections and Depth Supervision in Contrast-Enhanced CT Images), overlapping anatomy degrades algorithmic performance.

3D Volume (Ribs + Lungs + Nodule) 
       │
       ▼ (Forward Projection / X-Ray)
2D Image Plane (Rib Shadow overlaps Nodule) ──► Standard CNN fails (Confuses edges)

Digital Layer Separation via Generative Models

To prevent models from failing due to occlusion, a prominent line of research focuses on bone suppression—digitally removing the occluding structures before performing downstream diagnostic tasks.

Using image-to-image translation frameworks, networks can be trained to map a standard radiograph to a "soft-tissue-only" image. For instance, Jia Liang demonstrated in Bone Suppression on Chest Radiographs With Adversarial Learning that Generative Adversarial Networks (GANs) can effectively isolate and erase ribs and clavicles, revealing underlying focal pneumonia and subtle lesions.

To maximize efficiency, the authors of Image to Images Translation for Multi-Task Organ Segmentation and Bone Suppression in Chest X-Ray Radiography (Mohammad Eslami et al.) extended this concept to a multi-task framework. By modifying a single architecture to output both a bone-suppressed image and an organ segmentation mask simultaneously, they forced the network to learn shared representations of anatomy, preserving fine structural details that independent networks might erase.

Input CXR ──► [ Shared Encoder ] ──┬──► [ Bone Suppression Decoder ] ──► Soft-Tissue Image
                                   └──► [ Segmentation Decoder ]     ──► Clean Organ Mask

Disentanglement Learning via 3D Simulation

A core challenge of bone suppression is the scarcity of paired training data; you cannot easily take an X-ray of a patient with and without their ribs. To bypass this limitation, researchers leverage 3D computed tomography (CT) scans to simulate 2D projections.

Because a CT scan provides a full 3D volume, bones and soft tissues can be segmented perfectly in 3D space. Researchers can then digitally project these isolated volumes to generate perfectly paired 2D radiographs: one with bones, and one without.

In X-Ray In-Depth Decomposition: Revealing the Latent Structures, Shadi Albarqouni et al. used these simulated projections alongside physics-based constraints to train a model that decomposes a single 2D X-ray into its constituent depth layers.

Similarly, Luyi Han et al. introduced RSGAN in GAN-based disentanglement learning for chest X-ray rib suppression. Their framework explicitly disentangles anatomical structure from image contrast. By learning from CT-derived projections, the model learns to strip away the high-frequency edges of the ribcage while preserving the underlying vascular patterns and low-contrast lesions.


Multi-View and Multi-Scale Invariance

When a single projection is too heavily occluded to separate digitally, models must find alternative ways to resolve ambiguity, such as reasoning across multiple views or scales.

In mammography, dense glandular tissue frequently obscures masses. Yuhang Liu addressed this in Cross-View Correspondence Reasoning Based on Bipartite Graph Convolutional Network for Mammogram Mass Detection by modeling the geometric relationships between two different radiographic views of the same tissue. If a mass is occluded by dense tissue in the craniocaudal (CC) view, the bipartite graph network correlates features with the mediolateral oblique (MLO) view, resolving the spatial ambiguity.

For single-view systems, networks must rely on multi-scale feature fusion to maintain robustness. As shown by Mengran Zhou et al. in Pulmonary diseases accurate recognition using adaptive multiscale feature fusion in chest radiography, combining lightweight multiscale networks with deep backbones allows a model to capture broad contextual shapes (like an entire lung boundary) simultaneously with faint, localized pixel variations (like an early-stage infection hidden behind a rib).


Forcing Robustness via Synthetic Occlusion

Rather than trying to remove or bypass occlusion at inference time, another powerful paradigm involves forcing models to become invariant to occlusion during training. This is achieved by intentionally synthesizing "hard examples" where target features are buried behind dense structures.

In Disentangling Disease-related Representation from Obscure for Disease Prediction, Chu-ran Wang et al. utilized an alpha-blending technique to superimpose target lesions directly onto highly dense, obscuring backgrounds. Because the target and the background share identical visual properties (such as intensity and color), the model cannot rely on simple thresholding. It is forced to learn the intrinsic geometric properties of the lesion itself.

[ Clean Lesion Patch ] ──┐
                         ├──► [ Alpha Blending ] ──► Synthetic Hard Example (Occluded Lesion)
[ Dense Bone Background ] ─┘

This strategy of data augmentation through generation was further validated by the NODE21 challenge (Ecem Sogancioglu et al. in Nodule Detection and Generation on Chest X-Rays: NODE21 Challenge) and by Zhenrong Shen 0001 et al. in Image Synthesis with Disentangled Attributes for Chest X-Ray Nodule Augmentation and Detection. By generating synthetic nodules that mimic the exact texture and size of real pathologies and placing them directly beneath bone structures, they created training distributions that explicitly penalize models that fail under occlusion. As a result, when these models encounter real-world radiographs, they successfully "see through" the bones, isolating the underlying clinical signals.

Key Foundational Work

Automatic Detection Of Rib Contours in Chest Radiographs (1975) — contributed 2 concept elements to this direction.
Automatic Detection Of Rib Contours in Chest Radiographs (1975) — contributed 1 concept elements to this direction.
Tomography of the Temporal Bone (1972) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_anatomical_occlusion_5e77068f,
  title = {anatomical occlusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anatomical-occlusion/rdr_5e77068f90f14dc1bbb79a7f58cea0c0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_anatomical_occlusion_5e77068f_v1,
  title = {anatomical occlusion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anatomical-occlusion/rdr_5e77068f90f14dc1bbb79a7f58cea0c0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake