Standard image processing treats a picture as a flat grid of intensities. While intuitive, this representation fails when structures overlap. Imagine a bowl of spaghetti: in a 2D photograph, two strands crossing each other occupy the same pixel. To a computer, these strands become "glued" together, making it nearly impossible to track a single noodle from end to end. Remco Duits and colleagues describe this as the "crossing problem," a fundamental hurdle in medical imaging and computer vision.
The solution emerging across these papers is lifting: transforming a low-dimensional image into a higher-dimensional state space where overlapping structures become disentangled.
Disentangling the Crossing Problem
The most common form of lifting involves adding a dimension for orientation. In a 2D image, a pixel at $(x, y)$ can only tell you the brightness at that spot. In a "lifted" orientation score, that same point is expanded into a stack of values, each representing the presence of a structure at a specific angle $\theta$.
As Nicky J. van den Berg, Olga Mula, and Leanne Vis demonstrate in Connected Components on Lie Groups and Applications to Multi-Orientation Image Analysis, two blood vessels that cross in 2D actually pass "over" and "under" each other in the 3D space of $(x, y, \theta)$. By performing analysis in this lifted space, algorithms can identify connected components—like individual retinal vessels—without them merging at intersections.
This concept extends to 3D medical volumes. Michiel H. J. Janssen and Erik J Bekkers show in Design and Processing of Invertible Orientation Scores of 3D Images for Enhancement of Complex Vasculature that lifting 3D data into a 5D orientation score allows for the enhancement of complex vasculature that traditional 3D filters would simply blur together.
From Shortest Paths to Smooth Curves
Lifting isn't just about separating objects; it's about making optimization easier. Traditional "minimal path" models often take jagged shortcuts across corners because they only care about minimizing distance.
By lifting the problem into a space that includes orientation, researchers like Da Chen, Jean-Marie Mirebeau, and Laurent D. Cohen can penalize curvature. In A New Finsler Minimal Path Model with Curvature Penalization for Image Segmentation and Closed Contour Detection, they find paths that are not just short, but physically plausible. Similarly, Alexey Mashtakov and Yu. L. Sachkov use sub-Riemannian geometry on the projective line bundle to track vessels, significantly reducing the "cusps" or sharp U-turns that plague 2D models.
Lifting also transforms difficult, non-convex problems into manageable ones. Daniel Cremers, Thomas Pock, and Antonin Chambolle pioneered this in An algorithm for minimizing the Mumford-Shah functional. By lifting the image into a higher dimension, they created a convex relaxation of the Mumford-Shah functional, allowing for global optimization that doesn't get stuck in local minima.
Geometric Deep Learning on Lie Groups
Modern research is integrating these geometric insights into neural networks. Standard Convolutional Neural Networks (CNNs) are not naturally aware of rotations. Erik J Bekkers and Michaël Defferrard address this in ChebLieNet by lifting data into Lie groups. This ensures the network is equivariant: if an input image rotates, the network's internal representation rotates accordingly, leading to more robust learning.
Furthering this, Finn M. Sherry and Kristina Schaefer introduce Diffusion-Shock PDEs for Deep Learning on Position-Orientation Space. By embedding specialized partial differential equations (PDEs) into a geometric deep learning framework, they allow models to adaptively smooth or sharpen images while preserving the underlying geometry of crossing structures.
The Computational Bottleneck
The primary challenge of this direction is the "curse of dimensionality." Every paper, from Fethallah Benmansour to Daan Bon, notes that lifting data significantly increases memory and processing requirements.
- Lifting a 2D image into an orientation score creates a 3D volume.
- Lifting a 3D medical scan creates a 5D or 6D dataset.
- Adding scale (as in Upanshu Sharma's work on the similitude group $SIM(2)$) adds even more complexity.
While these methods provide unparalleled geometric clarity and mathematical rigor, the field's current frontier is finding ways to implement these high-dimensional transforms efficiently enough for real-time clinical and industrial use.