bovine-embryo-video-evaluation / docs /annotation_protocol.md
embryo-video-eval's picture
Update docs/annotation_protocol.md
61f04e0 verified

Annotation Protocol

Overview

This document describes the label structure and annotation process used in the BovEmbryo dataset. Each video is assigned an embryo type label and a final developmental outcome label. These annotations define the six-class classification task, the three-class outcome-only classification task, and the IVF ↔ SCNT embryo-origin generalization task.

Data Collection Context

Videos were collected using a MIRI Time-Lapse Incubator (Esco Medical) under single-embryo culture conditions. The incubator maintained a stable culture environment at 38.5°C with 5% CO2 and 5% O2. Brightfield imaging was used for non-invasive observation, with images automatically captured every 5 to 10 minutes. Videos were recorded at a resolution of 1280 × 720. Individual time-lapse sequences span approximately 183 to 200 hours of biological development, covering the full developmental window from the initial zygote stage through the terminal outcome on Day 7.

Each released file is an exported .avi time-lapse video. The digital playback duration is typically around 1–2 minutes and should not be interpreted as the biological development duration.

Annotation Unit

The annotation unit is one complete exported time-lapse video. Each video represents a single embryo's full developmental trajectory and receives:

  • one embryo type label
  • one final developmental outcome label

Labels are video-level labels. The dataset does not provide frame-level annotations, developmental-event timestamps, or pixel-level segmentation masks.

Annotators

Labels were assigned by domain experts in bovine embryology and assisted reproduction with professional experience in IVF and SCNT embryo outcome assessment. Annotation was performed by specialists familiar with standard bovine embryo development and outcome classification criteria used in reproductive biology practice. No crowdsourcing platform was used.

Annotation Method

Each annotator reviewed the complete exported time-lapse video and assigned an embryo type label and a developmental outcome label based on direct visual assessment of embryo morphology across the full recording. In cases of visual ambiguity, the final outcome observed at the end of the recording period was used as the definitive label.

Annotation Reliability

To assess annotation reliability, a subset of 44 videos was independently labeled by three trained experts. Fleiss’ κ was 0.45 for the three-class outcome label, indicating moderate agreement, and 0.77 for the binary success versus non-success label, indicating substantial agreement. Final labels for this subset were assigned by majority vote.

These agreement results indicate that coarse viability separation is more reliable than distinguishing fine-grained failure modes. Errors and model behavior should therefore be interpreted with the known ambiguity between degenerate and arrested embryos in mind.

Embryo Type Labels

The dataset contains two embryo types:

Embryo Type Description
IVF In vitro fertilization embryo
SCNT Somatic cell nuclear transfer embryo

The embryo type label is used directly in Task 1 and as the domain variable in Task 3.

Outcome Labels

The dataset contains three final developmental outcome labels:

Outcome Description
degenerate Embryo degeneration or nonviable development
arrested Embryo development stops or fails to continue progressing before successful completion
success Embryo reaches the successful developmental outcome defined for this dataset

These outcome labels are used in Task 2 and Task 3.

Six-Class Label Construction

Task 1 uses the combination of embryo type and final outcome.

Label ID Six-Class Label
0 IVF-degenerate
1 IVF-arrested
2 IVF-success
3 SCNT-degenerate
4 SCNT-arrested
5 SCNT-success

This task evaluates whether a model can classify both embryo type and developmental outcome from the video.

Three-Class Label Construction

Task 2 collapses IVF and SCNT into a shared outcome label space.

Label ID Three-Class Label
0 degenerate
1 arrested
2 success

This task evaluates whether a model can classify developmental outcome independent of embryo type.

Embryo-Origin Generalization Label Use

Task 3 uses embryo type as the domain variable.

The two domains are:

  • IVF
  • SCNT

The classification labels are:

  • degenerate
  • arrested
  • success

The two evaluation protocols are:

Protocol Training / Validation Domain Test Domain
IVF_to_SCNT IVF SCNT
SCNT_to_IVF SCNT IVF

This task evaluates whether models trained on one embryo type generalize to the other.

Annotation Ambiguity and Limitations

Some outcome categories may be visually similar, particularly degenerate and arrested embryos, especially in early developmental stages. Visual differences between IVF and SCNT embryos may also be subtle from video appearance alone. Model performance should be interpreted with these biological and visual ambiguities in mind.

The dataset provides video-level outcome labels only. It does not provide frame-level annotations, developmental-event timestamps, or pixel-level segmentation masks. The labels reflect terminal developmental outcome in the recorded setting and should not be interpreted as validated pregnancy, live-birth, or transfer-success labels.

Responsible Interpretation

The annotations are intended for non-commercial academic research and retrospective benchmark evaluation of biological video understanding models. They should not be used directly for clinical, veterinary, reproductive, or commercial decision-making without further validation, regulatory review, explicit authorization, and expert oversight.