Guided Conditional Diffusion Classifier (ConDiff) for Enhanced Prediction of Infection in Diabetic Foot Ulcers

TL;DR

The paper tackles automated infection detection in Diabetic Foot Ulcers from wound images. It introduces ConDiff, a Guided Conditional Diffusion Classifier that synthesizes conditionally guided images via CFG-DDIM and classifies infection by minimizing the $L_2$ distance in a learned embedding space, aided by a Triplet loss to reduce overfitting. On a refined DFU dataset, ConDiff achieves up to $83\%$ accuracy and an $F_1$-score of $0.858$, outperforming state-of-the-art baselines by at least $3\%$, albeit with higher inference time. This work advances generative-discriminative medical image analysis and suggests a path toward more accurate and interpretable infection detection in DFUs, with potential extensions to multi-modal data and other wound types.

Abstract

To detect infected wounds in Diabetic Foot Ulcers (DFUs) from photographs, preventing severe complications and amputations. Methods: This paper proposes the Guided Conditional Diffusion Classifier (ConDiff), a novel deep-learning infection detection model that combines guided image synthesis with a denoising diffusion model and distance-based classification. The process involves (1) generating guided conditional synthetic images by injecting Gaussian noise to a guide image, followed by denoising the noise-perturbed image through a reverse diffusion process, conditioned on infection status and (2) classifying infections based on the minimum Euclidean distance between synthesized images and the original guide image in embedding space. Results: ConDiff demonstrated superior performance with an accuracy of 83% and an F1-score of 0.858, outperforming state-of-the-art models by at least 3%. The use of a triplet loss function reduces overfitting in the distance-based classifier. Conclusions: ConDiff not only enhances diagnostic accuracy for DFU infections but also pioneers the use of generative discriminative models for detailed medical image analysis, offering a promising approach for improving patient outcomes.

Figures (12)

• Figure 1: Natural data augmentation of an original image with three different magnifications.
• Figure 2: Inference in the ConDiff Classier Framework. Input $x_0$ is perturbed by noise of strength $t_0$. The perturbed input $x_{t_0 T}$ is denoised through a reverse diffusion process to synthesize image $\hat{x}_0^{(y_i)}$ conditioned on label $y_i$. Infection classification is based on the minimum $L_2$ distance between $x_0$ and $\hat{x}_0^{(y_i)}$ in embedding space.
• Figure 3: Synthesizing conditional DFU images using ConDiff. A guide image $x_0$ is perturbed with Gaussian noises that are then removed progressively using a CFG-DDIM sampling technique (see Algorithm. \ref{['alg_cfg_sampling']}), conditioned on the infection status. This process gradually projects $x_0$ to guided synthetic images of conditions: $\hat{x}_0^{(y_1)}$ and $\hat{x}_0^{(y_2)}$.
• Figure 4: The learning accuracy trajectories of the ConDiff classifier and best-performing CNN-based (EfficientNet-B0) and ViT-based (EfficientFormer-L1) models, on the train and validation sets.
• Figure 5: Examples of incorrectly classified DFU images for infection by our ConDiff Classifier.