article.tex

\documentclass{article}

\usepackage{arxiv}

\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc}    % use 8-bit T1 fonts
\usepackage{hyperref}       % hyperlinks
\usepackage{url}            % simple URL typesetting
\usepackage{booktabs}       % professional-quality tables
\usepackage{multirow}       % multi-row cells in tables
\usepackage{amsfonts}       % blackboard math symbols
\usepackage{nicefrac}       % compact symbols for 1/2, etc.
\usepackage{microtype}      % microtypography
\usepackage{graphicx}
\usepackage{caption}
\usepackage{amsmath}
\usepackage[numbers]{natbib}

\title{DFU Segmentation and False Positive Reduction: A Clinical Decision Support System using EfficientNet and Vision Transformers}

\author{
  Alexandre dos Santos Pereira \\
  \href{mailto:alexandre.202111286@unilasalle.edu.br}{alexandre.202111286@unilasalle.edu.br} \\
  \href{https://github.com/Nucleo-IA-Unilasalle/dfu-segmentation-net}{github.com/Nucleo-IA-Unilasalle/dfu-segmentation-net} \\
}

\begin{document}
\maketitle

\begin{abstract}
Diabetic Foot Ulcers (DFU) are a serious complication of diabetes that requires timely and accurate monitoring. While automated segmentation models can assist in wound delineation, they often suffer from high false positive rates when deployed in real-world uncontrolled environments where other skin lesions with wounds can interfere with the output. This paper presents a robust two-stage deep learning framework designed as a clinical decision support tool. The first stage utilizes an EfficientNet-B4 U-Net for high-sensitivity binary semantic segmentation. The second stage employs a Vision Transformer (ViT) based Verification Classifier to filter false positives. By integrating the predicted segmentation mask and segmentation percentage as inputs to the ViT, the system effectively distinguishes true wounds from background noise. Our approach achieves a specificity of 95.2\% with the ViT verifier, significantly outperforming baseline CNN methods, and provides a practical workflow for validating image quality before clinical review.
\end{abstract}

\section{Introduction}
Diabetic Foot Ulcers (DFU) represent a significant burden on healthcare systems and patient quality of life. Accurate measurement and monitoring of wound area are critical for assessing healing progress. Manual segmentation is time-consuming and subject to inter-observer variability. Deep learning offers the potential for automated, objective assessment.

However, deploying segmentation models in clinical settings faces a major challenge: \textbf{False Positives}. Models trained on curated wound datasets often over-segment or misclassify non-wound objects (such as socks, furniture, or healthy skin) as ulcers when applied to images taken in uncontrolled environments. In a clinical workflow, excessive false alarms can lead to "alarm fatigue" or distrust in the system.

We propose a clinical decision support system that acts as a "gatekeeper." Instead of replacing the clinician, the system verifies the presence of a wound with high specificity. If the system detects a non-wound or has low certainty, it prompts the user to retake the photo, ensuring that only high-quality, relevant data reaches the medical professional.

\section{Methodology}

\subsection{Dataset Construction}
To build a robust system, we curated a diverse set of data for both training and "hard negative" mining:
For positive samples representing wounds, we utilized the Wound Segmentation Dataset \cite{wang2020wound} (used for both segmentation and classification tasks) and the Leprosy Chronic Wound Dataset as our primary training data (used as a test set). Conversely, to train the verification classifier with negative samples (non-wounds), we sourced images from MiniImageNet for general objects. We also included specific "hard negatives" from Skin Cancer and Skin Disease datasets, such as HAM10000 \cite{tschandl2018ham10000}, to help the model distinguish DFU from other dermatological conditions.

For the wound classification task, we utilized a dataset of \textbf{1,019} manually annotated wound images with multi-label annotations for exudate quantity and tissue type called \textbf{PI-Wound Classification Dataset}. This dataset was created by anotatting the images from the Wound Segmentation Dataset \cite{wang2020wound}. Table \ref{tab:dataset_distribution} presents the distribution of samples across categories.
\begin{table}[h]
  \caption{"PI-Wound Classification Dataset" Category Distribution}
  \centering
  \begin{tabular}{lcc}
    \toprule
    Category & Count & Percentage \\\midrule
    None & 610 & 59.86\% \\
    Low & 179 & 17.57\% \\
    High & 139 & 13.64\% \\
    Medium & 91 & 8.93\% \\
    \midrule
    \multicolumn{3}{l}{\textit{Tissue Categories}} \\
    \midrule
    Granulação & 751 & 73.70\% \\
    Necrotic & 126 & 12.37\% \\
    Esfacelo & 105 & 10.30\% \\
    Epitelial & 37 & 3.63\% \\
    \midrule
    \multicolumn{3}{l}{\textit{Total Samples}} \\
    \midrule
    Total & 1,019 & 100.00\% \\
    \bottomrule
  \end{tabular}
  \label{tab:dataset_distribution}
\end{table}

\subsection{Stage 1: Segmentation Network}
The core segmentation model is an \textbf{EfficientNet-B4 U-Net}.
The network employs an EfficientNet-B4 encoder \cite{tan2019efficientnet} pretrained on ImageNet, selected for its balance of accuracy and computational efficiency. This is paired with a custom U-Net style decoder \cite{ronneberger2015unet}, where skip connections bridge the encoder and decoder layers to preserve spatial resolution critical for boundary delineation. To handle class imbalance between small wound areas and large backgrounds, the model is trained using Dice Loss and optimized with Adam (LR=0.001). Prioritizing sensitivity, this stage achieves a Dice Coefficient of approximately 89.3\%, accepting a higher false positive rate to ensure no potential wounds are missed.

\subsection{Stage 2: Verification Classifier}
To address the false positives generated by the high-sensitivity segmentation model, we implement a secondary binary classifier model. We tested multiple models and input strategies, but the best performing model was a Vision Transformer (ViT) model with soft-labels as input. Performance with methods such as CNNs and XGBoost were tested and resulted in a very similar but slightly worse performance.

\subsubsection{Verification Classifier Training Data}
Training a robust false positive filter requires both diverse positive (wound) examples and challenging negative (non-wound) examples. We assembled a comprehensive dataset from multiple sources:

\textbf{Positive Samples (Wound Images):} We utilized two primary wound datasets. The \textbf{Wound Segmentation Dataset} \cite{wang2020wound} is a publicly available dataset containing chronic wound images with pixel-level segmentation annotations. We used the standard train/test split provided by the dataset authors, employing training images for classifier training and test images for evaluation. Additionally, we incorporated the \textbf{Leprosy Chronic Wound Dataset (CO2Wounds-v2)}, a dataset of chronic wound images from leprosy patients that provides additional wound diversity. We applied an 80/20 split, using the majority for training and reserving 20\% as an independent test set to evaluate generalization to unseen wound types.

\textbf{Negative Samples (Non-Wound Images):} We assembled four distinct datasets representing common sources of false positives in clinical deployment:

\begin{itemize}
  \item \textbf{MiniImageNet}: A subset of ImageNet containing 60,000 color images across 100 classes of everyday objects (furniture, animals, household items) that may appear in uncontrolled photography environments.
  \item \textbf{Skin Cancer Dataset}: A collection of skin cancer images used to challenge the classifier with dermatological imagery exhibiting similar color profiles and textures to wounds but representing benign or malignant skin lesions rather than ulcers.
  \item \textbf{Skin Disease Dataset}: A comprehensive dermatological dataset containing images of various skin conditions (eczema, psoriasis, fungal infections, etc.). These images serve as "hard negatives" since they may visually resemble wound tissue but require different clinical management.
  \item \textbf{HAM10000} \cite{tschandl2018ham10000}: The Human Against Machine dataset containing 10,015 dermatoscopic images of pigmented skin lesions across 7 diagnostic categories (melanoma, nevus, basal cell carcinoma, etc.). These high-quality clinical images represent the most challenging false positive cases.
\end{itemize}

The training procedure creates a balanced dataset by randomly sampling equal numbers of wound and non-wound images, ensuring the classifier does not develop a bias toward either class.

\subsubsection{Input Strategy}
We hypothesize that the segmentation map itself contains valuable context (e.g., shape, location) that can aid classification.
The classifier inputs consist of a 4-channel image (created by concatenating the original RGB image with the predicted binary mask from Stage 1) and a scalar segmentation percentage representing the proportion of the image predicted as a wound. This additional scalar helps filter noise from extremely small regions or massive failures where the entire image is segmented.

\subsubsection{Wound Characteristic Classification}
Beyond wound detection, we also evaluated multi-class classification for wound characteristics: exudate quantity (none, low, medium, high) and tissue type (granulation, necrotic, slough, epithelial). We compared four feature extraction approaches:

\textbf{Handcrafted Features (XGBoost-Simple):} Color statistics (RGB/HSV mean, std, median, percentiles), texture features via Local Binary Patterns (LBP) with 26-bin histograms and gradient magnitude statistics, and shape features from contour analysis (area, perimeter, circularity, aspect ratio, extent, solidity). Total: $\sim$55 features.

\textbf{Radiomics Features (XGBoost-Radiomics):} A focused radiomics feature set extracted from masked wound regions, including RGB color statistics (mean, std, median for each channel), grayscale statistics (mean, std, entropy, Laplacian variance, edge density), GLCM texture properties (contrast, homogeneity, energy, correlation), LBP texture features (mean, std, uniformity), and peri-wound transition features (15 features analyzing wound-to-healthy-skin boundaries). Total: $\sim$36 features.

\textbf{Combined Features (XGBoost-Combined):} Handcrafted features augmented with radiomics-inspired features including Gray-Level Co-occurrence Matrix (GLCM) properties (contrast, homogeneity, energy, correlation), peri-wound region analysis (15 features capturing wound-to-healthy-skin transitions), and additional LBP statistics. Total: $\sim$77 features.

\textbf{Vision Transformer (ViT):} End-to-end learning using a pretrained ViT-Base model fine-tuned on masked wound regions with soft-label training using BCEWithLogitsLoss.

All models were trained with soft labels derived from multi-annotator consensus, using sample weighting based on label confidence for XGBoost models.

    \subsubsection{Classification Head}
    We modified the first convolutional projection layer (\texttt{conv\_proj}) to accept four channels, initializing the mask channel weights to zero to allow gradual learning without disrupting the pretrained RGB features. The resulting 768-dimensional feature vector from the ViT encoder is then concatenated with the segmentation percentage scalar. Finally, a multi-layer perceptron (MLP) classification head maps these combined features to a binary output distinguishing wounds from non-wounds.


\subsection{Evaluation Metrics}
We employ standard classification and segmentation metrics to evaluate our two-stage system. For binary classification tasks (Stage 2), we use the following metrics based on true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN):

Sensitivity ($\text{Sensitivity} = \frac{TP}{TP + FN}$) measures the proportion of actual wounds correctly identified, where high sensitivity ensures minimal missed wounds. Specificity ($\text{Specificity} = \frac{TN}{TN + FP}$) tracks the proportion of non-wounds correctly rejected, minimizing false alarms. Precision ($\text{Precision} = \frac{TP}{TP + FP}$) indicates the reliability of positive predictions by measuring the proportion of predicted wounds that are actual wounds. Finally, the F1 Score ($\text{F1} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$) provides a balanced harmonic mean of precision and recall.

For segmentation evaluation (Stage 1), we use:
The Dice Coefficient ($\text{Dice} = \frac{2|X \cap Y|}{|X| + |Y|}$) measures the spatial overlap between the predicted mask $X$ and the ground truth mask $Y$, ranging from 0 (no overlap) to 1 (perfect overlap). Consequently, Dice Loss ($\mathcal{L}_{\text{Dice}} = 1 - \text{Dice}$) serves as the training objective for the segmentation network to effectively handle class imbalance.

\section{Clinical Workflow and Logic}
The system is designed to function as follows:
The process begins when a clinician or patient captures an image of the foot, which is then processed by the EfficientNet-B4 U-Net to generate a candidate mask. Subsequently, the ViT model analyzes both the image and the mask for verification. If the system verifies the input as a wound, it presents the segmentation result and area measurements to the clinician. However, if the system detects a non-wound or has low certainty, it triggers a feedback prompt advising the user to check lighting, remove background clutter (such as socks), and retake the photo.

\section{Results}

We evaluated the system on independent test sets, including specific "Specificity Tests" using skin lesion datasets to mimic hard false positives.

\subsection{Segmentation Model Training Comparison}
To justify our choice of EfficientNet-B4 as the encoder backbone, we compared training metrics between a vanilla U-Net and our EfficientNet-B4 U-Net. Figure \ref{fig:unet_comparison} presents the training and validation curves for both architectures.

\begin{figure}[h]
  \centering
  \includegraphics[width=\textwidth]{unet_vs_effunet_metrics.png}
  \caption{Training comparison between vanilla U-Net and EfficientNet-B4 U-Net. The EfficientNet backbone demonstrates faster convergence and superior final performance across all metrics.}
  \label{fig:unet_comparison}
\end{figure}

Table \ref{tab:segmentation_comparison} summarizes the final performance metrics for both architectures on the validation set.

\begin{table}[h]
  \caption{Segmentation Model Comparison: Vanilla U-Net vs EfficientNet-B4 U-Net (Best Values)}
  \centering
  \begin{tabular}{lccc}
    \toprule
    Model & Best Dice Score & Lowest Val Loss & Best Val IoU \\
    \midrule
    Vanilla U-Net & 0.76 & 0.240 & 0.62 \\
    EfficientNet-B4 U-Net & \textbf{0.89} & \textbf{0.110} & \textbf{0.81} \\
    \bottomrule
  \end{tabular}
  \label{tab:segmentation_comparison}
\end{table}

The EfficientNet-B4 backbone provides several advantages over the vanilla U-Net: (1) pretrained ImageNet weights enable better feature extraction, (2) compound scaling ensures optimal encoder capacity, and (3) improved convergence stability during training. These benefits translate to higher Dice scores and more consistent wound boundary delineation.

\subsection{Segmentation Model False Positive Analysis}
To quantify the false positive challenges, we evaluated the Stage 1 EfficientNet-B4 U-Net on three non-wound datasets: MiniImageNet (general objects), Skin Cancer (HAM10000 and others), and a Skin Disease dataset. Since these datasets contain no actual wounds, any segmentation output represents a false positive. We define an image as a false positive if more than a given percentage threshold of the output mask pixels are classified as wound. Table \ref{tab:fp_analysis} presents the false positive rates at three thresholds: $>0\%$ (any segmentation), $>1\%$, and $>5\%$ of image area. Lower values indicate better performance.

\begin{table}[h]
  \caption{Segmentation Model False Positive Rate on Non-Wound Datasets}
  \centering
  \begin{tabular}{llll}
    \toprule
    Dataset & FP Rate ($>0\%$) & FP Rate ($>1\%$) & FP Rate ($>5\%$) \\
    \midrule
    MiniImageNet & 40.58\% & 12.62\% & 4.28\% \\
    Skin Cancer & 85.04\% & 68.24\% & 37.15\% \\
    Skin Disease & 60.34\% & 21.60\% & 7.80\% \\
    \bottomrule
  \end{tabular}
  \label{tab:fp_analysis}
\end{table}

The segmentation model shows a high propensity for false positives, particularly on the Skin Cancer dataset where 85\% of images produced some segmentation output. Even when requiring substantial segmentation ($>5\%$ of image area), the false positive rate remains problematically high at 37.15\% for Skin Cancer. These results highlight the need for a secondary verification stage to filter spurious detections before clinical review.




\begin{figure}[h]
  \centering
  \includegraphics[width=\textwidth]{model_predictions.png}
  \caption{Qualitative examples of segmentation results. Each column shows the original image, ground truth mask, predicted probability map, and final binary prediction with IoU and Dice scores.}
  \label{fig:model_predictions}
\end{figure}

   \begin{table}[h]
    \caption{Comparison of Verification Models (CNN vs ViT)}
     \centering
     \begin{tabular}{lllll}
       \toprule
       Model & Sensitivity & Specificity & F1 Score & Precision \\
       \midrule
       CNN Baseline & 57.28\% & 94.20\% & 67.03\% & 80.79\% \\
       ViT Verifier & 56.34\% & 95.20\% & \textbf{67.23\%} & 83.33\% \\
       \bottomrule
     \end{tabular}
     \label{tab:results}
   \end{table}
   
\subsection{Wound Characteristic Classification Results}
Table \ref{tab:wound_classification} presents the comparison of different approaches for wound characteristic classification on the test set.

\begin{table}[h]
  \caption{Wound Characteristic Classification: Test Set Performance Comparison}
  \centering
  \begin{tabular}{l|ccc|ccc}
    \toprule
    & \multicolumn{3}{c|}{Exudate Classification} & \multicolumn{3}{c}{Tissue Classification} \\
    Model & Accuracy & Macro F1 & Weighted F1 & Accuracy & Macro F1 & Weighted F1 \\
    \midrule
    XGBoost-Simple & 70\% & 0.36 & 0.62 & 75\% & 0.38 & 0.68 \\
    XGBoost-Radiomics & 69\% & 0.33 & 0.60 & 75\% & 0.36 & 0.66 \\
    XGBoost-Combined & 68\% & 0.33 & 0.60 & 75\% & 0.39 & 0.68 \\
    ViT (Ours) & \textbf{75\%} & \textbf{0.61} & \textbf{0.75} & \textbf{86\%} & \textbf{0.63} & \textbf{0.84} \\
    \bottomrule
  \end{tabular}
  \label{tab:wound_classification}
\end{table}

Table \ref{tab:wound_classification_detail} provides detailed per-class F1 scores for the best performing model (ViT).

\begin{table}[h]
  \caption{ViT Per-Class F1 Scores on Test Set}
  \centering
  \begin{tabular}{lc|lc}
    \toprule
    \multicolumn{2}{c|}{Exudate Quantity} & \multicolumn{2}{c}{Tissue Type} \\
    Class & F1 Score & Class & F1 Score \\
    \midrule
    None & 0.87 & Granulation & 0.91 \\
    Low & 0.44 & Necrotic & 0.85 \\
    Medium & 0.44 & Slough & 0.76 \\
    High & 0.69 & Epithelial & 0.00 \\
    \bottomrule
  \end{tabular}
  \label{tab:wound_classification_detail}
\end{table}

The ViT model significantly outperforms handcrafted feature approaches, achieving 7\% higher accuracy on exudate classification and 11\% higher on tissue classification. Notably, the macro F1 score improvement is substantial (0.61 vs 0.36 for exudate, 0.63 vs 0.38 for tissue), indicating better performance on minority classes. Both the combined radiomics features (XGBoost-Combined) and the focused radiomics-only approach (XGBoost-Radiomics) performed similarly to the simple handcrafted features, with macro F1 scores around 0.33-0.39. This suggests that end-to-end learning with ViT captures more discriminative representations than manually engineered features for this task, regardless of whether radiomics features are included or not.

   \subsection{Analysis}
   The ViT model outperforms the CNN baseline in filtering false positives. The ViT maintains high specificity (95.2\%) and precision (83.33\%) while achieving a sensitivity of 56.34\%, making it suitable for a clinical support tool where both trust and detection capability are paramount.

   When optimizing for false positive reduction, \textbf{specificity} is the most critical metric to maximize, as it directly measures the proportion of non-wounds correctly rejected. High specificity ensures that the system minimizes false alarms, which is essential for maintaining clinician trust and preventing alarm fatigue. Precision serves as a complementary metric, indicating the reliability of positive predictions; a high precision means that when the system flags an image as containing a wound, it is likely correct. In clinical decision support systems where excessive false positives can lead to workflow disruption and reduced confidence in the tool, prioritizing specificity over sensitivity represents the appropriate trade-off.


\section{Conclusion}
We presented a DFU segmentation framework that effectively mitigates false positives through a novel two-stage approach. By leveraging an EfficientNet-B4 U-Net \cite{tan2019efficientnet,ronneberger2015unet} for segmentation and a Vision Transformer \cite{dosovitskiy2020image} that consumes both image and mask data for verification, we achieved high specificity suitable for real-world deployment. This system empowers clinicians by automating wound measurement while actively filtering poor-quality or irrelevant images, streamlining the documentation workflow in diabetic foot care.

\begin{thebibliography}{99}

\bibitem{tan2019efficientnet}
Tan, M., \& Le, Q. (2019).
\newblock EfficientNet: Rethinking model scaling for convolutional neural networks.
\newblock In {\em International Conference on Machine Learning} (pp. 6105--6114).
\newblock PMLR.

\bibitem{ronneberger2015unet}
Ronneberger, O., Fischer, P., \& Brox, T. (2015).
\newblock U-net: Convolutional networks for biomedical image segmentation.
\newblock In {\em International Conference on Medical Image Computing and Computer-Assisted Intervention} (pp. 234--241).
\newblock Springer.

\bibitem{wang2020wound}
Wang, C., Anisuzzaman, D.~M., Williamson, V., Dhar, M.~K., Rostami, B., Niezgoda, J., Gopalakrishnan, S., \& Yu, Z. (2020).
\newblock Fully automatic wound segmentation with deep convolutional neural networks.
\newblock {\em Scientific Reports}, 10, 21897.
\newblock \url{https://doi.org/10.1038/s41598-020-78799-w}

\bibitem{tschandl2018ham10000}
Tschandl, P., Rosendahl, C., \& Kittler, H. (2018).
\newblock The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.
\newblock {\em Scientific Data}, 5, 180161.
\newblock \url{https://doi.org/10.1038/sdata.2018.161}

\bibitem{dosovitskiy2020image}
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., \& Houlsby, N. (2020).
\newblock An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.
\newblock {\em arXiv preprint arXiv:2010.11929}.

\end{thebibliography}

\end{document}