[Paper Review] Translating Clinical Delineation of Diabetic Foot Ulcers into Machine Interpretable Segmentation

Connah Kendrick, Bill Cassidy+7 more|arXiv (Cornell University)|Apr 22, 2022

Diabetic Foot Ulcer Assessment and Management17 citations

TL;DR

This paper introduces DFUC2022, the largest clinical segmentation dataset for diabetic foot ulcers (DFUs) with 4,000 images annotated by clinicians. It demonstrates that preprocessing clinician-delineated contours with image processing to create smoothed masks significantly improves deep learning segmentation performance, achieving a Dice score of 0.7446—surpassing the best baseline (0.5708) and the DFUC2022 challenge winner (0.7287).

ABSTRACT

Diabetic foot ulcer is a severe condition that requires close monitoring and management. For training machine learning methods to auto-delineate the ulcer, clinical staff must provide ground truth annotations. In this paper, we propose a new diabetic foot ulcers dataset, namely DFUC2022, the largest segmentation dataset where ulcer regions were manually delineated by clinicians. We assess whether the clinical delineations are machine interpretable by deep learning networks or if image processing refined contour should be used. By providing benchmark results using a selection of popular deep learning algorithms, we draw new insights into the limitations of DFU wound delineation and report on the associated issues. This paper provides some observations on baseline models to facilitate DFUC2022 Challenge in conjunction with MICCAI 2022. The leaderboard will be ranked by Dice score, where the best FCN-based method is 0.5708 and DeepLabv3+ achieved the best score of 0.6277. This paper demonstrates that image processing using refined contour as ground truth can provide better agreement with machine predicted results. DFUC2022 will be released on the 27th April 2022.

Motivation & Objective

• To create the largest clinical segmentation dataset for diabetic foot ulcers (DFUC2022) with ground truth annotations by clinicians.
• To investigate whether raw clinician-delineated contours are suitable as ground truth for deep learning models or if image processing refinement improves performance.
• To benchmark state-of-the-art deep learning segmentation networks on the DFUC2022 dataset and identify key performance limitations.
• To propose and validate a new training strategy that improves segmentation accuracy by modifying the FCN32 VGG architecture to address data distribution, noise, and region similarity issues.
• To demonstrate that smoothed masks from image processing lead to better alignment with model predictions than jagged clinical contours.

Proposed method

• Collected and curated 4,000 DFU images (2,000 train, 2,000 test) with clinical delineations from expert clinicians.
• Applied image processing techniques to refine raw, jagged clinician-drawn contours into smoothed, machine-interpretable masks for improved ground truth representation.
• Trained and evaluated a range of deep learning segmentation models (e.g., DeepLabv3+, FCN32 VGG) on both raw and smoothed masks to compare performance.
• Conducted ablation studies to isolate the impact of data distribution bias, background noise, and texture similarity on model performance.
• Modified the FCN32 VGG network architecture to reduce excessive down-sampling and improve feature learning on small, low-contrast DFU regions.
• Used explainable AI techniques (e.g., class activation maps) to analyze model attention and verify that networks focus on ulcer regions during prediction.

Experimental results

Research questions

• RQ1 Can clinical delineations of DFUs be directly used as ground truth for deep learning segmentation, or does image processing refinement improve model performance?
• RQ2 What are the primary challenges in training deep learning models on DFU segmentation, and how do data distribution, noise, and texture similarity affect performance?
• RQ3 Does using smoothed masks derived from clinician contours lead to better alignment between model predictions and clinical annotations?
• RQ4 Can architectural modifications to standard segmentation networks (e.g., FCN32 VGG) improve performance on small, low-contrast DFU regions without post-processing?
• RQ5 How does the performance of the proposed method compare to state-of-the-art models on the DFUC2022 benchmark, particularly in terms of Dice score and robustness to noise and small lesions?

Key findings

• The DFUC2022 dataset is the largest available DFU segmentation dataset to date, with 2,000 training and 2,000 test images annotated by clinicians.
• Image processing refinement of clinician-drawn contours into smoothed masks significantly improved segmentation performance, with a Dice score of 0.7446 compared to 0.5708 for the best baseline model.
• The proposed modified FCN32 VGG network achieved a Dice score of 0.7446, outperforming the first-place model in the DFUC2022 challenge leaderboard (0.7287).
• Baseline models struggled with small DFU regions due to excessive down-sampling and data distribution bias, where over 95% of images had DFU regions below 5% of total image area.
• Model performance was significantly degraded by background noise and texture similarity between DFU, periwound, and healthy skin, leading to misclassification and overfitting to background.
• Rapid up-sampling in standard networks caused pixelation and false positive regions, which were mitigated by architectural modifications that improved feature map resolution and localization.