Capsule Review

The question

Capsule endoscopy produces long image sequences. Selecting the highest-scoring frames may repeatedly capture one long positive region while missing shorter events. This project asks whether model-guided selection can cover more distinct blood-positive annotation runs than uniform sampling when both select the same number of frames.

What I built

I built a local research viewer with sequence navigation, human notes, bookmarks and intervals, imported model scores, deterministic review queues and verifiable evidence exports. Unselected images remain accessible; model scores stay separate from human findings.

The experimental work includes compact-CNN training and evaluation, original-image verification, fixed-budget policy comparisons and independent numerical replay. Later studies tested training diversity and short-event ranking across excluded examinations.

Viewer and selection code · Tests

Locked comparison

The Galar 11–20 study used 163,161 source-verified images from ten examinations, including 1,742 blood-positive annotated images in 154 contiguous annotation runs. Six examinations contained positive runs; recall is undefined for the other four.

1. 01Verify original images and chronology
2. 02Select the same exact K frames per policy
3. 03Measure frame retrieval and distinct-run coverage

The compact blood CNN and four deterministic policies were fixed before the corrected evaluation outcomes were inspected. Policies included uniform midpoint sampling, unrestricted highest-score ranking, score per chronological bin and score with temporal spacing. The prespecified primary compared score-per-bin with uniform selection at 10% of each examination’s annotated frames.

The mean paired difference was −1.27 percentage points across the six positive-run examinations: two improved, two declined and two tied. The 25% secondary comparison was +9.51 points, with substantial variation between examinations; it does not reverse the negative primary result.

Frozen evaluation contract · Preserved public aggregate

Frames are not runs

A policy can retrieve many positive frames from relatively few runs. The highest-score policy illustrates this failure at the same 10% budget:

Here, an annotation run means consecutive blood-positive labels at the dataset’s native annotation stride. It is not an independently adjudicated lesion. Pooled run counts and equally weighted examination means answer different questions.

Limitations

This is a retrospective annotation-coverage study, not clinical validation. Examination IDs have not been verified as unique patients, and the sampled annotations are not complete video-event ground truth. No clinician reading-time savings or lesion sensitivity were measured. The evaluated 11–20 examinations are already inspected; the dataset authors’ reserved 61–80 outcomes remain unopened.

Try the software

The local demonstration contains 24 synthetic images and generated, non-AI scores. From a checkout of the repository:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e .
capsule-review serve

Open http://127.0.0.1:8765 and choose Open test sequence. The walkthrough includes Windows setup, queue comparison and evidence export.

To check selected saved public aggregates without medical images or model training, run python scripts/project_status.py. See the verification guide for software tests and reproduction boundaries. Full medical-image experiments require private inputs; they cannot be reproduced from the public aggregate alone.

The repository contains no medical images, individual predictions or trained pathology weights. Source is publicly visible without an explicit software licence.

Current research direction

Later development did not establish reliable cross-examination improvement. A frozen MobileNet ranking probe reached 49.08% pair ordering versus 72.29% for the compact CNN on 69 matched boundary comparisons. This small, already-inspected development task is distinct from whole-examination run recall.

The next useful step is qualified expert interpretation of short events and ambiguous labels before another model comparison. A private 137-image blinded pilot is prepared but has not been reviewed or distributed. Reviewer participation, data-use requirements and adjudication must be settled first.

Current status · Scientific synthesis · Representation study