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

This study addresses extravasation, a critical issue requiring prompt detection for effective management to avoid severe complications. This work harnesses the capability of zero-shot capabilities from pre-trained vision transformer models-specifically, GroundingDINO and Segment Anything Model (SAM) for segment human skin regions, and Contrastive Language-Image Pretraining (CLIP) for extracting rich features from these regions using its frozen image encoder. Our methodology applies linear probe techniques to feature vectors obtained from CLIP using few-shot instances. The results demonstrate accurate classification of extravasation severities with training using only 64 instances per class, achieving average F1_macro scores of 74.08% for GroundingDINO-CLIP. This marks an improvement result to the previous study which utilized dual U-Nets for skin and lesion segmentation alongside DenseNet-121 models for classification with training using 975 instances. Notably, this approach increases previous F1_macro scores by 3.27% in mild extravasation cases. This research advances fine-grained extravasation classification, specifically in early detection, achieved through few-shot models in a unique context.