NICU-Care Dataset

Citation Author(s):
Yue
Xin
Shanghai Jiao Tong University
Submitted by:
Yue Xin
Last updated:
Fri, 04/11/2025 - 22:11
DOI:
10.21227/kwxh-v002
Data Format:
License:
0
0 ratings - Please login to submit your rating.

Abstract 

NICU-Care is a high-quality video dataset designed to support visual recognition tasks in Neonatal Intensive Care Unit (NICU) scenarios, including nursing action recognition, object detection, and semantic segmentation. It was constructed in a standardized simulated NICU environment, capturing multi-view RGB videos of professional nurses performing six types of routine caregiving procedures on simulated infants. The dataset provides fine-grained temporal annotations and pixel-level segmentation masks for key objects like nurse hands, medical tools, and infant body parts. It replicates real-world clinical workflows and adopts a structured annotation system to support diverse research objectives. The dataset's design, annotation methodology, and task adaptability were rigorously validated using mainstream deep learning models. A unified data format and storage scheme ensure structured management and reusability. The public release of NICU-Care is expected to facilitate automated nursing assessment, intelligent monitoring, and data-driven training, with future extensions planned to incorporate richer annotations and multimodal sensory information to further increase its scientific and practical value.

Instructions: 

The NICU nursing dataset has a hierarchical file structure. At the top level, directories are named after action categories (e.g., the Change Diaper/ folder contains all instances of the diaper change action). Within each action folder, the data is further separated by camera view, usually subfolders like Top/ and Right/, since each clinical program is shot from two different angles. This structure (Action → View) allows for easy access to all videos for a particular action, with the option to focus on a specific viewpoint if needed. The name of each video file contains a wealth of information about the recording. We have adopted a uniform naming convention: ActionCode_View_NurseID_Instance.mp4. For example, a file named CD_Rview_A_1.mp4 can be interpreted as follows: CD denotes the “change diaper” action, the Rview indicates capture from the right camera view, A is the anonymous identifier of the nurse (caregiver) performing the operation, and 1 indicates that this is the first recorded instance of a diaper change operation performed by nurse A. The following is an example of a diaper change operation. Similarly, the top view recording of the second diaper change by the same nurse might be labeled CD_Tview_A_2.mp4, and a different action code would be used for the feeding action (FB for bottle feeding). By embedding category, viewpoint, subject, and session information in the file names, we ensure that each file is self-descriptive and unique. This naming scheme and folder organization greatly facilitates data retrieval. One can query files by action type or viewpoint at a glance, while also laying the groundwork for program access.

Documentation

AttachmentSize
File README - 副本 (2).txt1.63 KB