In standard computer vision, we often rely on generic geometric features like corners or blobs to align images. However, in the human body, these features are often unreliable. Tissues deform, disease alters local textures, and different imaging modalities (like MRI and Ultrasound) "see" the same organ in fundamentally different ways.
To solve this, researchers are moving away from generic geometry toward anatomically stable landmarks. These are specific biological structures—vessel junctions, cartilage boundaries, or urethral points—that remain consistent across different patients, imaging technologies, and disease states. By treating these structures as "anatomical anchors," we can achieve registration and diagnostic accuracy that generic algorithms cannot match.
Bridging the Modality Gap
One of the hardest problems in medical imaging is registration: aligning a high-resolution pre-operative scan (like MRI) with a noisy, real-time scan (like Ultrasound). Because the physics of these sensors are different, a "corner" in an MRI might not exist in an ultrasound.
Xinrui Song and colleagues addressed this in Distance Map Supervised Landmark Localization for MR-TRUS Registration. Instead of looking for raw pixels, they trained networks to predict distance maps of stable landmarks like the urethral entry and exit points. These points serve as a common geometric basis that bridges the gap between magnetic resonance and acoustic signals.
Similarly, Zhongliang Jiang found that costal cartilage is a superior landmark for Autonomous US-CT Registration. While bone is often used, cartilage has a distinct acoustic signature and a visible pleural line that makes it a more reliable anchor for navigating robotic ultrasound through narrow intercostal windows.
Invariance Across Disease and Demographics
A robust landmark must be "invariant"—it should look the same whether the patient is young or old, healthy or sick.
In retinal imaging, disease can drastically change the appearance of the fundus, making standard keypoint matching fail. Yu Wang developed SuperJunction, which focuses specifically on vessel bifurcations and crossovers. These structural intersections are domain-specific anchors that persist even when disease progression alters the surrounding tissue's texture.
This principle of demographic invariance is also seen in cardiac imaging. Xiaoping Liu demonstrated that the central septum cross remains stable across diverse patient ages and sizes, allowing for the automatic acquisition of the four-chamber view in 3D echocardiography without needing to process the entire, highly variable volume of the heart.
Accounting for Physical Displacement
Sometimes, the challenge isn't just finding a landmark, but choosing one that doesn't move when the anatomy fails. In severe rotator cuff tears, muscles can retract (pull away) from their original position. If a clinician uses a standard lateral "Y-view" to assess the muscle, they might see an empty space and misclassify a reparable tear as irreparable atrophy.
Guoyin Zhang proposed a solution in A new section of shoulder magnetic resonance imaging used to assess fat infiltration. By shifting the imaging plane medially to a stable landmark—the suprascapular notch (the "r-view")—the scan captures the true muscle belly regardless of how much the distal end has retracted. This medialized view provides significantly higher inter-observer reliability because it anchors the diagnosis to a part of the anatomy that stays put.
From Points to Edges
While many landmarks are points, some are more robust as higher-order structures. In the complex, reflective environment of a colonoscopy, point features are often obscured by specular reflections. Danail Stoyanov and the SegCol Challenge team highlight that anatomical fold edges represent consistent depth discontinuities. These edges are more stable for 3D reconstruction and navigation than traditional points because they are defined by the fundamental geometry of the organ's interior rather than transient lighting conditions.
By identifying and prioritizing these stable anatomical features—whether they are vessel junctions, cartilage signatures, or specific medial planes—researchers are creating AI systems that understand the "ground truth" of human anatomy, leading to more precise surgeries and more accurate diagnoses.