LISP-Net: Clinician-Guided AI for Context-Aware Segmentation

LISP-Net is a deep learning model that propagates a 2D ROI across distant slices using a single example as context. Given one annotated slice, it predicts the segmentation on other slices — adapting to your intention, anatomy, and style.

It feels natural: draw on one slice, then let the model carry your intent forward. Interact with the model if it comes off track.

Crucially, LISP-Net runs entirely client-side in your browser — powered by ONNX Runtime Web with WebGPU acceleration. No server, and no uploads. Your data never leaves your device.

1. Usage

LISP-Net is keyboard-driven — there is no menu or toolbar. All commands are single-key presses.

Prerequisites

First, use the ROI Tool to draw a single 2D ROI on one slice.

Step 1 — Memorize a Context Slice

With your ROI drawn on the reference slice, press M. The first time, LISP-Net loads the ONNX model into the browser (may take a few seconds), then immediately saves the current image and mask as the prompt context — the pair the model will use to understand and propagate your intent. A notification confirms when ready.

The model needs only one annotated slice. On subsequent uses, pressing M simply updates the memorized context.

Step 2 — Choose the Normalization Mode

Press T to toggle between CT and MRI normalization. This is critical for correct results:

Mode	Use for
CT	CT images (HU units). Applies fixed windowing: `[-1000, 1000]` clipped, shifted by `+15`, scaled by `1/160`
MRI	MRI and everything else. Foreground percentile clipping (p0.5–p99.5) followed by per-volume z-score normalization

The current mode is shown in a notification and persists across browser sessions.

Step 3 — Predict on a Distant Slice

Scroll to any other slice and press N. LISP-Net predicts the segmentation on that slice using the memorized context. The result is written into your active ROI immediately.

Batch Prediction — Hold N + Scroll

Hold down N and scroll through a range of slices. When you release N, LISP-Net will calculate predictions for every slice you passed through.

Performance depends on image size, number of slices, and your hardware.

Step 4 — Correct and Refine

If a prediction is not perfect:

Edit the ROI on that slice using the ROI tool
Press M again to save the corrected slice as a new context
Continue predicting on further slices

The model benefits from updated context — each correction improves subsequent predictions.

2. Expert Tips

Feed Good Predictions as New Context

Even if a prediction is already good, you can press M on that slice to use it as fresh context. As you move further from the original slice, appearance may drift — updating context keeps the model aligned to the local anatomy.

Switching Slicing Orientation

LISP-Net respects the current slicing dimension. If you switch between axial, coronal, or sagittal views, be aware that the memorized prompt was saved in a specific orientation. The model will warn you if you try to predict across slices in a different orientation.

Cancelling

If a batch prediction is in progress and you want to stop it, press N again — the current run will be cancelled.

3. Keyboard Shortcuts Summary

Key	Action
M	Memorize current slice as context (first press initializes the model if not loaded)
N	Predict on current slice (press & hold + scroll + release for batch prediction across a range)
T	Toggle CT / MRI normalization mode