Extravasation Screening and Severity Prediction from Skin Lesion Image using Deep Neural Networks

Area of Science:

• Medical Imaging and Artificial Intelligence
• Dermatology and Clinical Diagnostics

Background:

• Extravasation, leakage of intravenous medication, causes severe soft tissue injury and necrosis.
• Delayed treatment can necessitate surgical debridement, skin grafting, or amputation.
• Accurate and timely assessment of extravasation severity is crucial for effective management.

Purpose of the Study:

• To develop a smartphone application for predicting extravasation severity using skin images.
• To leverage Deep Neural Network (DNN) architectures for automated extravasation screening.

Main Methods:

• Utilized U-Net and DenseNet-121 DNN architectures for skin and lesion segmentation.
• Employed DNNs for classifying extravasation severity (asymptomatic, mild, moderate, severe).
• Proposed a rule-based system to enhance multi-class classification accuracy for moderate-to-severe cases.

Main Results:

• Achieved 77.78% sensitivity and 90.24% specificity for distinguishing asymptomatic from abnormal cases.
• Mild extravasation classification yielded the highest F1-score (0.8049).
• Moderate and severe extravasation classifications achieved F1-scores of 0.6250 and 0.6429, respectively.

Conclusions:

• A novel and feasible DNN approach for screening extravasation from skin images was proposed.
• Mobile DNN applications show strong potential for clinical use, especially in low-resource settings.
• The application can serve as a valuable tool for monitoring extravasation and optimizing clinical workflows.