NICUface: Robust Neonatal Face Detection in Complex NICU Scenes
Bibliographic record
Abstract
The development of non-contact patient monitoring applications for the neonatal intensive care unit (NICU) is an active research area, particularly in facial video analysis. Recent studies have used facial video data to estimate vital signs, assess pain from facial expression, differentiate sleep-wake status, detect jaundice, and in face recognition. These applications depend on an accurate definition of the patient’s face as a region of interest (ROI). Most studies have required manual ROI definition, while others have leveraged automated face detectors developed for adult patients, without systematic validation for the neonatal population. To overcome these issues, this paper first evaluates the state-of-the-art in face detection in the NICU setting. Finding that such methods often fail in complex NICU environments, we demonstrate how fine-tuning can increase neonatal face detector robustness, resulting in our NICUface models. A large and diverse neonatal dataset was gathered from actual patients admitted to the NICU across three studies and gold standard face annotations were completed. In comparison to state-of-the-art face detectors, our NICUface models address NICU-specific challenges such as ongoing clinical intervention, phototherapy lighting, occlusions from hospital equipment, etc. These analyses culminate in the creation of robust NICUface detectors with improvements on our most challenging neonatal dataset of +36.14, +35.86, and +32.19 in AP30, AP50, and mAP respectively, relative to state-of-the-art CE-CLM, MTCNN, img2pose, RetinaFace, and YOLO5Face models. Face orientation estimation is also addressed, leading to an accuracy of 99.45%. Fine-tuned NICUface models, gold-standard face annotation data, and the face orientation estimation method are also released here.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".