An ingenious deep learning approach for pressure injury depth evaluation with limited data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Wound Care

Background:

• Deep learning (DL) models for pressure injury assessment from wound images are gaining traction.
• Acquiring sufficient supervised data for DL model training is time-consuming.
• Developing high-performing DL models with limited supervised data is crucial.

Purpose of the Study:

• To develop and compare the performance of three deep learning models for assessing pressure injury depth.
• To evaluate the effectiveness of different artificial intelligence (AI) classification strategies.
• To determine the optimal approach for pressure injury depth classification using limited data.

Main Methods:

• A retrospective observational study analyzed 414 sacral pressure injury images (d0-D4) using DL.
• Three convolutional neural network models were developed: Categorical, Binary, and Combined classification.
• Model performance was evaluated using metrics like F1-score and precision.

Main Results:

• The Combined classification model demonstrated superior performance with an F1-score of 0.868.
• The Binary classification model effectively differentiated between d0 and d1-D4 stages (F1-score, 0.895).
• The Categorical model excelled in classifying advanced stages (D3/D4) but struggled with early stages (d1/d2 misclassified as d0).

Conclusions:

• The Combined classification model achieved high performance without requiring additional supervised data.
• This superior performance is attributed to a hybrid approach using a Binary model for initial assessment and a Categorical model for subsequent stages.
• Strategic deployment of AI classification methods tailored to specific wound characteristics can significantly enhance model performance in pressure injury assessment.