Protocol for metadata and image collection at diabetic foot ulcer clinics: enabling research in wound analytics and deep learning

Abstract

Background: The escalating impact of diabetes and its complications, including diabetic foot ulcers (DFUs), presents global challenges in quality of life, economics, and resources, affecting around half a billion people. DFU healing is hindered by hyperglycemia-related issues and diverse diabetes-related physiological changes, necessitating ongoing personalized care. Artificial intelligence and clinical research strive to address these challenges by facilitating early detection and efficient treatments despite resource constraints. This study establishes a standardized framework for DFU data collection, introducing a dedicated case report form, a comprehensive dataset named Zivot with patient population clinical feature breakdowns and a baseline for DFU detection using this dataset and a UNet architecture.

Results: Following this protocol, we created the Zivot dataset consisting of 269 patients with active DFUs, and about 3700 RGB images and corresponding thermal and depth maps for the DFUs. The effectiveness of collecting a consistent and clean dataset was demonstrated using a bounding box prediction deep learning network that was constructed with EfficientNet as the feature extractor and UNet architecture. The network was trained on the Zivot dataset, and the evaluation metrics showed promising values of 0.79 and 0.86 for F1-score and mAP segmentation metrics.

Conclusions: This work and the Zivot database offer a foundation for further exploration of holistic and multimodal approaches to DFU research.

Keywords: Case report form; Deep learning; Depth; Diabetic foot ulcer; RGB; Thermal.

Fig. 1

Illustration of image aspect of the Zivot dataset. The images were taken in two sessions, before and after debridement, with each session consisting of three different angles relative to the ulcer surface. The top row of the sample dataset shows the RGB images, whilst the next two rows illustrate the corresponding depth and thermal maps. In the maps, the colour red represents closer distances from the camera or higher temperatures in the depth and thermal maps, respectively. The colour blue indicates further distances from the camera or lower temperatures. Any pixel with no available value is shown in black, indicating missing data

Fig. 2

ROC curve analysis. ROC curve showcases the performance of the setup, indicating its ability to achieve high true-positive rates whilst maintaining low false-positive rates. The average AUC across the threefold is 0.98, with a standard deviation of 0.0082

Fig. 3

Versatile detection capabilities of the trained network. The red boxes represent the actual locations of wounds, whilst the green boxes show the predicted locations by the model. The following scenarios demonstrate the capabilities of the bounding box detection model. A Bounding box detection of a visible wound with surrounding box, confidence level, and bounding box area in pixels. B Early detection of a DFU, even with partial skin coverage, before debridement. C Successful DFU detection in challenging conditions like a bright background with a post-debridement ulcer. D detection of multiple ulcers in a single image. E Accurate detection of diverse skin tones, including dark skin. F Network specificity by correctly excluding a wound on the ankle from DFU detection. G and H illustrate failed cases where the second wound on the left foot is missed in G, and red nail polish is mistakenly marked as DFUs in H

Fig. 4

Camera setup and positioning. Left: the depth and thermal cameras were securely and adjacently positioned within a dedicated compartment. The RGB lenses were aligned and placed 3.3 cm from each other. The stereo and IR-generated depth, and thermal maps are superimposed on the RGB images for each camera. Right: the fixed apparatus was affixed to the rear of a laptop, enabling simultaneous image capture initiation from the laptop

Fig. 5

DL DFU detection architecture. A combination of EfficientNetb3 and UNet was used and trained with the Zivot dataset. This model has been shown to be a high-performing detection model for DFUs [46]