WORK IN PROGRESS · READER STUDY / B / 02
Two routes into the same OSA paper. Explore the moving illustration, inspect its modules, then test the mathematical intuition. Nothing in these simulations is a clinical prediction.
01 — The system
A visual walkthrough of a frame's spatial features and the memory carried across time. Select any module to see its role.
Generated geometry and texture, not medical footage, ground truth or a measured segmentation result.
The ResNet-50 backbone turns each ultrasound frame into feature maps. A feature map is a grid of numerical measurements, not the original pixels.
APFE decomposes each feature into a local mean and positive/negative contrasts. Learned convolution branches and an adaptive gate combine the cues; this is not perfect speckle removal.
OSU updates a matrix state with the new frame, then approximately projects it toward orthogonal columns. The state also flows from one time step to the next.
The decoder turns the tracked features into a segmentation. The animation here is a teaching example, not an actual decoder prediction.
Follow the arrows to see the sequence; the lower loop shows the state returning from the previous frame.
02 — The research question
Echocardiograms change shape rapidly and contain severe speckle. OSA tackles the spatial ambiguity with APFE and the temporal memory problem with OSU.
The ventricle changes appearance throughout the cardiac cycle.
Speckle and low contrast can obscure its boundary.
Unconstrained recurrent state directions may lose independence over time.
03 — A little mathematics
Move the angle between two unit vectors. The geometry below explains two terms you will encounter in OSU without requiring linear algebra first.
The arrows point in different directions, but the matrix is not perfectly balanced.
Imagine drawing a circle on a rubber sheet and stretching it with a matrix. The circle becomes an ellipse. Singular values are how much its longest and shortest axes were stretched. The smallest tells us whether a direction has nearly disappeared.
Divide the biggest stretch by the smallest. A value of 1 means equally strong directions; a very large number means some directions are much harder to recover. When the smallest singular value is zero, the condition number is infinite.
An ideal orthogonal state has perpendicular unit columns, so its singular values are all 1. The actual paper uses a finite-step Newton–Schulz approximation after state updates—not the perfect 2D correction shown here.
The arrows and numbers are an exact calculation on an invented 2×2 matrix. They do not represent OSA measurements or a clinical result.
04 — Experiment hands-on
The same APFE and OSU teaching experiments appear in both versions. Change one parameter at a time and compare the visual response.
INTERACTIVE WALKTHROUGH
Follow a frame through OSA: first separate complementary feature cues, then stabilize the evolving memory. Move the controls to explore the underlying mathematics.
A local average captures slow intensity variation. Positive and negative deviations retain complementary contrasts around that local field.
X = M + X+ − X−
The decomposition is algebraically exact, but it does not perfectly isolate anatomy: high-frequency speckle can remain in both residuals. In the real network, learned convolutions and an adaptive gate fuse these channels.
Both sides receive the same sequence of gated updates. Only the right side projects each intermediate state back to an orthogonal basis.
S⊤S = I
This toy example uses the paper's gated recurrence (Eq. 1) and an ideal 2D polar projection. Actual OSU uses a scaled, finite-step Newton–Schulz approximation; these bars are computed from toy matrices, not measured model states.
05 — Back to the paper
The actual OSA paper includes the network diagram, method derivations, and evaluations on CAMUS and EchoNet-Dynamic. Our interactive figures illustrate principles; they are not reproduced measurements.
Read the original evidence ↗