Redefining the Classification of Extravasation Severity Using CLIP Linear Probe with Few-shot Instances.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Extravasation requires prompt detection to prevent severe complications.
• Previous methods for extravasation classification were data-intensive and less accurate.

Purpose of the Study:

• To develop a few-shot learning approach for fine-grained extravasation classification using pre-trained vision transformer models.
• To improve the accuracy and efficiency of early extravasation detection.

Main Methods:

• Utilized GroundingDINO and Segment Anything Model (SAM) for human skin segmentation.
• Employed Contrastive Language-Image Pretraining (CLIP) for feature extraction from skin regions.
• Applied linear probing with few-shot instances for classification.

Main Results:

• Achieved average F1_macro scores of 74.08% for GroundingDINO-CLIP with only 64 instances per class.
• Demonstrated a 3.27% increase in F1_macro scores for mild extravasation cases compared to previous studies.
• Showcased superior performance over dual U-Nets and DenseNet-121 models with significantly less training data.

Conclusions:

• Few-shot learning with pre-trained vision transformers offers a highly effective method for extravasation classification.
• This approach advances early and fine-grained detection of extravasation, particularly in medical contexts.
• The developed methodology provides a more efficient and accurate alternative to existing extravasation detection systems.