Retinal Layer Segmentation

Overview

Retinal Layer Segmentation runs on OCT volume for segmenting retina. It runs a deep-learning model on every B-scan, detects the retinal layer boundaries, reconstructs them into smooth curves, and builds a 3-D view of the layers across the volume. Everything runs locally. Your data never leave the machine.

The extension opens its own window, titled RetinalLayerSegmentation.

What you can do

• Load OCT and OCTA volumes, load any previously saved segmentation.
• Segment the retina with one click, on the GPU if available.
• Review the detected boundaries, show or hide individual layers, and correct any boundary by hand.
• Interpolate the segmentation across sparse B-scans to fill the whole volume.
• Inspect the result as B-scans, en-face projections and a 3-D surface.
• Send the curves and masks back to the MedICS workspace, or save them to files.

Requirements

• MedICS, installed and running.
• GPU is required for accelate AI retinal layer segmentation.

Install

Install the extension from inside MedICS, the Settings-> Extension.

Quick start

The Retinal Layer Segmentation window

The window has two areas. The left panel holds everything you operate; the right side shows four linked views of the data.

Data source

The top of the left panel loads input data. It has two tabs.

File tab

Choose a file for each input with its folder button, or type the path, then click Load Data to read them together.

Array orientation matters. When a .mat file is loaded, its volume must be stored as nframe × nrow × ncol — B-frames along the first axis, rows (depth) second, and A-scans (columns) last. Transpose the array before loading if the axes are in another order, or the frames will be interpreted incorrectly.

• OCTA checkbox — uncheck it to load an OCT-only dataset. Also will hide loaded OCTA data from the Left B-frame Panel
• setting… — opens Load Data Settings, where you can review, assign and manage the loader functions used for each data type.
• Load Data — reads everything selected at once.

Load Data Settings

The setting… button on the File tab opens Load Data Settings. This is where you choose which loader function reads each kind of file. Open it when a file is not recognised, or to teach the extension a custom format.

How to use it:

1. Select a loader function in the list on the left — its source appears in the preview below.
2. Click OCT >>, OCTA >> or Seg >> to assign the selected function to that input.
3. Type the file patterns the loader should handle in its Filters box.

Right-clicking a function in the list offers Delete, which does the same as Remove.

Save to WS tab

Transfer what is currently loaded back into the MedICS workspace under the names shown. Edit a name if you want a different one, then click the >> button in that row.

Drag and drop

Drop supported OCT, OCTA or segmentation files anywhere on the window. Files are matched by their name and routed to the right input automatically; you can also drag a variable out of the MedICS  Workspace Variables panel and drop it on the window to load it. This extension also support drop a image sequence (same image type) as OCT / OCTA volume data.

Frame & resolution

Scan Resolution

Upate data resolution opens the Scan Resolution dialog. It converts the physical size of the scan into a per-pixel resolution so that measurements and the 3-D view are to scale.

Enter the two required values and the lateral resolution updates live. Click OK to write the new resolution to the loaded data.

The axial resolution cannot be derived from the image — it is fixed by the OCT device. The help button opens a reminder to consult the device specifications.

Region of interest

Restrict processing to a vertical band of each B-scan — this speeds segmentation up and avoids the empty space above and below the retina.

Segmentation controls

These controls drive the AI model and the manual refinement. They sit below the data and preprocessing groups.

Run AI Segmentation

Runs the deep-learning model on every B-scan and fills in the layer boundaries. If a GPU device is available you are asked which one to use; otherwise it runs on the CPU. Progress and any errors are reported in the status bar.

Device

Chooses how the model runs. AUTO picks the best available device; named entries are the GPUs found on the machine; CPU always works. If a GPU fails, the extension falls back to the CPU automatically.

Corrector

Selects the manual editing tool. All correcting happens in the Right B-Frame (the OCT panel). Choose a tool from the Corrector dropdown — None for review only, or one of the three editing tools below.

First decide how the target boundary is chosen:

The selection mode is set from the Right B-Frame context menu (see B-Frame context menus). The correctors are available in the Fast scan view; in Slow scan view they are hidden.

Liner

Draws a straight segment over the boundary — the quickest way to correct a short span.

1. Hover over the boundary to focus it; it thickens while the other boundaries dim.
2. Click the left mouse button at the start of the segment, then drag to the end point. A straight line follows the pointer and replaces the curve between the two points. During correction, click left mouse button to reset the start point of the segment.
3. Double click left mouse botton to stop and apply correction.

Only the dragged span changes; the rest of the boundary stays as it was.

LiveWire

Traces the strongest image edge between two points — best when the boundary is clearly visible in the image but awkward to draw by hand.

1. Hover over the boundary to focus it.
2. Click the left mouse button at the start point, then move the pointer along the boundary. The tool keeps computing the lowest-cost path along the image edges and redraws the curve as you move. During correction, click left mouse button to reset the start point.
3. Double click left mouse botton to stop and apply correction.

Because it follows image edges, LiveWire can be pulled off-course by speckle or fluid; re-trace a shorter span if that happens.

Anchors

Edits the boundary point by point, by dragging its control points (the small square markers). Use it for precise, local adjustments.

1. Hover over the boundary to focus it — its control points appear.
2. Move mouse  to a control point to select it; the cursor marker snaps to it.
3. Hold the left button and drag to move the point. The curve refits through the anchors as you drag.
4. While dragging, scroll the mouse wheel up to add the neighbouring points on each side to the selection (they move together), and scroll down to drop them again.

Right-click the panel while Anchors is active for these entries:

Interpolation

Save

Saves the current segmentation to a file (curves and, optionally, the fluid mask). To send the result to the MedICS workspace instead, use the Save to WS tab.

Layer boundaries

The Layer boundaries group lists every boundary the model can detect. Tick a box to show and export that boundary; use the small dropdown next to it to choose how it is drawn (dashed or solid line). Select All toggles every boundary at once.

Only the boundaries that default to On are produced by the AI model. The entries that default to Off — PVD, GCL/IPL, ELM, IZ/RPE and SAT/HAL — are listed for completeness but are not implemented in the auto-segmentation model. Ticking one of them does not detect or create a curve; leave them off unless you plan to draw that boundary by hand with the corrector.

A few boundaries are listed for completeness but are not produced by the current model; enabling them does not add a curve. The remaining boundaries are detected automatically.

The views

The right side is split into four linked views. Moving the frame — or picking a point — keeps them in sync.

View toolbar

Each view has a small toolbar. Hover a button to see what it does.

B-Frame context menus

Right-click inside a B-Frame panel to open its context menu. The two panels differ: the Left B-Frame (OCT / OCTA overlay) focuses on display and export, while the Right B-Frame (the editable OCT panel) also offers the corrector tools.

Left B-Frame (OCT / OCTA overlay)

Right B-Frame (OCT)

The image-export entries (Save image, Save stack, Save video) are available on the Left B-Frame; on the Right B-Frame the menu is dedicated to editing and workspace transfer.

Status bar

The bar along the bottom reports where the pointer is and what the data is.

One-off messages — loading, saving, warnings — also appear here.

Keyboard shortcuts

The window has no menu bar; these are the keys you can use while working.

Mouse & interaction

Menus & context menus

The window itself has no menu bar; every action lives on a control in the left panel or on a view toolbar.

Saving & exporting

GPU acceleration & troubleshooting

Segmentation is far faster on a supported GPU. It uses an ONNX Runtime model with an optional NVIDIA CUDA provider. GPU acceleration needs the driver, the ONNX Runtime build and the CUDA/cuDNN libraries to match each other.

To check the installed build, run python -c "import onnxruntime as ort; print(ort.get_available_providers())" — you want CUDAExecutionProvider listed.

Notes & limitations

• Only boundaries produced by the model are drawn; a few listed boundaries are placeholders.
• Automatic CPU fallback means a too-old GPU still yields a result, just more slowly.
• Sparse volumes must be interpolated to obtain a dense segmentation.
• The segmentation is computed locally; no image data is uploaded.