FUTransUNet: Hybrid Transformer-U-Net for DFU Segmentation

• The paper introduces a DFU-specific hybrid model that combines a CNN encoder, Vision Transformer bottleneck, and U-Net decoder with Grad-CAM for visual explanation.
• It demonstrates strong segmentation performance with high Dice (≈0.875) and IoU metrics by leveraging both local feature extraction and global context modeling.
• The architecture addresses challenges of heterogeneous ulcer appearance and complex clinical images, potentially enhancing objective wound analysis and treatment planning.

FUTransUNet is a hybrid Transformer-U-Net segmentation model introduced for automatic diabetic foot ulcer (DFU) segmentation from clinical photographs. In the paper “FUTransUNet-GradCAM: A Hybrid Transformer-U-Net with Self-Attention and Explainable Visualizations for Foot Ulcer Segmentation,” the model combines a U-Net-style convolutional encoder, a Vision Transformer bottleneck, and a skip-connected decoder, and couples this segmentation pipeline with Grad-CAM visualizations for interpretability (Asare et al., 4 Aug 2025). The name is sometimes used in the paper as “FUTransUNet” and sometimes simply as “TransUNet,” but the intended contribution is a DFU-specific hybrid architecture rather than a generic label for all Transformer-U-Net systems. The term is also easy to confuse with adjacent models in the literature, especially FocalUNETR, whose official name is distinct and whose application domain is CT prostate segmentation rather than foot-ulcer photographs (Li et al., 2022).

1. Definition and nomenclature

FUTransUNet denotes a medical image segmentation framework for DFU delineation in clinical RGB photographs. Its defining structural pattern is an encoder → transformer bottleneck → decoder pipeline with U-Net skip connections, together with Grad-CAM-based visual explanation of model predictions (Asare et al., 4 Aug 2025). Within the source paper, the terminology is not perfectly stable: the model is sometimes described as FUTransUNet and sometimes simply as TransUNet. The architectural intent, however, is explicit: a hybrid Transformer-U-Net tailored to ulcer segmentation.

A frequent point of confusion is the relation between FUTransUNet and other transformer-augmented U-shaped networks. FocalUNETR, for example, is officially “Focal U-NEt TRansformers,” but it is a separate model designed for CT-based prostate segmentation, using a hierarchical focal transformer encoder and an auxiliary contour regression task rather than a CNN encoder with a ViT bottleneck (Li et al., 2022). Likewise, AFTer-UNet is a 2.5D axial-fusion Transformer-U-Net for volumetric medical segmentation, and TF-UNet is an MLP-augmented U-Net for optical fiber imaging rather than a transformer architecture (Yan et al., 2021, Xu et al., 2 Feb 2026).

This naming ambiguity matters because “FUTransUNet” can be read informally as “Fusion” or “Foot Ulcer” Transformer U-Net. In the present literature record, the most direct and official match is the DFU model of 2025 (Asare et al., 4 Aug 2025).

2. Clinical target and design rationale

The clinical problem addressed by FUTransUNet is segmentation of diabetic foot ulcers in clinical photographs. Accurate delineation is presented as important for wound area measurement, longitudinal healing assessment, treatment planning, and more objective clinical documentation (Asare et al., 4 Aug 2025). Manual segmentation is described as slow, subjective, and prone to inter-observer variability.

The segmentation task is difficult because DFUs exhibit heterogeneous appearance, irregular boundaries, variable size and shape, subtle contrast with surrounding tissue, and cluttered or complex backgrounds. The motivating claim of the architecture is that standard CNNs such as U-Net provide strong localization but are constrained by the local receptive field of convolution, and therefore are weaker at modeling long-range spatial dependencies. In this framing, ulcer delineation depends simultaneously on local texture and edge evidence and on whole-image context.

FUTransUNet addresses this by distributing representational roles across three subsystems. The CNN encoder is used for local feature extraction, especially texture, edges, and boundary cues. The Vision Transformer bottleneck is inserted to model global relationships between distant regions through self-attention. The decoder, with U-Net skip connections, restores spatial detail and precise localization. This division of labor is the central design principle of the model (Asare et al., 4 Aug 2025).

A plausible implication is that FUTransUNet belongs to the class of bottleneck-hybrid architectures rather than full-transformer U-Nets. It does not replace the entire encoder with transformer blocks; instead, it preserves a conventional convolutional feature pyramid and injects global contextual reasoning at the deepest stage.

3. Architecture

The model operates on RGB clinical photographs with original size 512 × 512 × 3, resized to 256 × 256 × 3 for training and inference. The output is a single-channel probability map, producing one ulcer probability per pixel. Final binary mask generation uses a threshold of 0.5, with ulcer pixels set to 255 and background pixels set to 0 (Asare et al., 4 Aug 2025).

The encoder is a four-stage U-Net-style CNN. Each encoder block contains two 2D convolution layers with kernel size 3 × 3 and same padding, followed by batch normalization and ReLU activation. After each block, 2 × 2 max pooling halves the spatial resolution. The number of filters increases across stages as 32, 64, 128, and 256. Intermediate feature maps are stored as $c_1, c_2, c_3, c_4$, and the deepest pooled feature map $p_4$ is fed into the transformer bottleneck (Asare et al., 4 Aug 2025).

The transformer bottleneck is a Vision Transformer module. The paper states $H=W=256$ and patch size $P=16$, yielding

$N = \frac{H \times W}{P \times P}$

and therefore

$N = \frac{256 \times 256}{16 \times 16} = 256.$

Thus the transformer processes 256 patch tokens. Each patch is flattened and passed through a linear projection to obtain a fixed-dimensional embedding, and learnable positional encodings are added before the sequence enters the transformer encoder. The bottleneck contains six sequential transformer encoder layers with 8 attention heads, layer normalization, feed-forward network (FFN / MLP), and GELU activation (Asare et al., 4 Aug 2025).

The self-attention mechanism is described using the standard multi-head formulation

$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V,$$

where $Q$, $K$, and $V$ are the query, key, and value projections and $p_4$0 is the key dimensionality. In this model, the transformer serves as a global reasoning block between encoder and decoder rather than as a full backbone (Asare et al., 4 Aug 2025).

The decoder mirrors the encoder and restores segmentation resolution through four decoder levels. Each stage performs 2 × 2 up-sampling, either by transposed convolution or nearest-neighbor interpolation as stated in the paper, concatenates the result with the corresponding encoder skip feature $p_4$1, and applies a convolutional refinement block of the same style as the encoder. The final segmentation head is a 1 × 1 convolution followed by sigmoid activation (Asare et al., 4 Aug 2025).

End to end, the feature flow can be summarized as follows: input image $p_4$2 convolutional encoder $p_4$3 transformer tokenization and self-attention $p_4$4 reshape back to feature map $p_4$5 skip-connected decoder $p_4$6 sigmoid probability map $p_4$7 thresholded ulcer mask. The architecture is therefore a canonical hybrid in which local and global representations are fused at decoding time.

4. Data, preprocessing, and optimization

Experiments are reported on the Foot Ulcer Segmentation Challenge (FUSeg) dataset, described as a public benchmark under the MICCAI framework. The dataset contains over 1,200 high-resolution clinical photographs, collected over a two-year period from hundreds of patients and de-identified according to HIPAA guidelines. The split reported in the paper is 810 training images with expert-annotated masks, 200 validation image-mask pairs, and 200 unlabeled test images reserved for final evaluation or submission (Asare et al., 4 Aug 2025).

Preprocessing is limited but explicit. Original images of 512 × 512 × 3 are resized to 256 × 256 × 3 to reduce computational complexity and GPU memory use while preserving enough anatomical detail. Predictions are thresholded at 0.5, and binary outputs are encoded as 255 for wound and 0 for background. The paper notes strong class imbalance in the masks, with background pixels dominating ulcer pixels (Asare et al., 4 Aug 2025).

A custom FootUlcerDataGenerator is used for efficient loading, preprocessing of image-mask pairs, and batching. The batch size is 16. Shuffling is enabled for training and disabled for validation. No explicit augmentation scheme is reported; the provided text does not specify operations such as rotation, flipping, scaling, color jitter, or elastic deformation (Asare et al., 4 Aug 2025).

Optimization uses Adam with initial learning rate 0.001 and Binary Cross-Entropy (BCE) as the training loss. Monitored metrics are accuracy, Dice coefficient, and IoU. Training runs for a maximum of 50 epochs with early stopping if validation loss does not improve for 10 consecutive epochs. The callback stack consists of ModelCheckpoint, ReduceLROnPlateau with factor 0.5 after 10 epochs without validation-loss improvement, and EarlyStopping with patience 10 and restoration of the best weights (Asare et al., 4 Aug 2025).

The paper does not report the exact GPU model, CPU, RAM, framework version, CUDA version, or operating system. This omission is notable because resizing is explicitly motivated by GPU memory reduction, but no computational profile is given.

5. Objective functions, reported performance, and Grad-CAM

The optimization objective is Binary Cross-Entropy. The paper does not print its explicit equation, but the provided standard form is

$p_4$8

Evaluation uses Dice, IoU, and pixel accuracy. The provided standard forms are

$p_4$9

$H=W=256$0

and

$H=W=256$1

with numerical stabilization constant

$H=W=256$2

The reported training and validation outcomes are as follows (Asare et al., 4 Aug 2025):

Metric	Training	Validation
Accuracy	0.9980	0.9971
IoU	0.7672	0.7780
Dice	0.8679	0.8751
Loss	0.0053	0.009045

The paper interprets the close training and validation values as evidence of good generalization and absence of obvious overfitting. It also notes that validation Dice and IoU are slightly higher than the corresponding training values. This suggests a relatively stable optimization trajectory on the reported split, although the paper does not provide external validation or statistical significance analysis (Asare et al., 4 Aug 2025).

A central auxiliary contribution is interpretability via Grad-CAM. The method is applied to the last convolutional layer of the network, not directly to transformer attention maps. The standard Grad-CAM formulation provided in the source material is

$H=W=256$3

followed by

$H=W=256$4

Here $H=W=256$5 denotes the $H=W=256$6-th feature map, $H=W=256$7 the gradient-derived importance weight, and $H=W=256$8 the number of spatial locations. The visual output is a coarse localization map overlaid on the input image, with red/yellow regions indicating higher influence. The paper states that these heatmaps align with anatomically and clinically relevant ulcer regions (Asare et al., 4 Aug 2025).

An important misconception to avoid is treating Grad-CAM in this work as an explanation of transformer attention itself. The reported implementation targets the final convolutional feature stage near prediction, so the explainability layer is CNN-centric even though the overall architecture is hybrid.

6. Position within the Transformer-U-Net literature

FUTransUNet sits within a broader family of medical image models that combine U-Net-style multiscale decoding with a mechanism for global context. Its closest conceptual relatives are bottleneck-hybrid Transformer-U-Nets, but the exact fusion strategy differs across models.

AFTer-UNet, for example, is a 2.5D hybrid model that uses a 2D CNN encoder, an Axial Fusion Transformer at the bottleneck, and a 2D CNN decoder with U-Net skip connections. Its defining contribution is sequential fusion of inter-slice and intra-slice attention, intended to recover volumetric context without full 3D attention cost (Yan et al., 2021). Relative to FUTransUNet, AFTer-UNet is more explicitly volumetric and slice-aware, whereas FUTransUNet is a 2D photographic segmentation model with a conventional ViT bottleneck.

FocalUNETR belongs to a different branch of the lineage. It replaces a CNN encoder with a hierarchical focal transformer encoder with patch merging and adds an auxiliary boundary-induced contour regression head for CT prostate segmentation. The paper explicitly distinguishes its official name from informal interpretations such as “Focal U-Net Transformer,” and does not introduce “FUTransUNet” as an alias (Li et al., 2022). Relative to FUTransUNet, the main differences are the modality, the encoder design, and the use of multitask boundary supervision.

TF-UNet is more distant architecturally. It is a physics-inspired reconstruction network for tapered optical fiber imaging that augments skip connections with hierarchical grouped-MLP fusion. It is relevant only at the level of design space: it demonstrates that non-local interaction can be inserted into a U-Net without self-attention, using global spatial mixing in skip pathways rather than transformer blocks (Xu et al., 2 Feb 2026).

These comparisons place FUTransUNet in the subclass of U-Net hybrids that preserve convolutional encoders, introduce transformer reasoning at the bottleneck, and rely on skip-connected decoding for localization. The model is therefore closer in spirit to classical TransUNet-style designs than to full hierarchical transformer encoders or non-attention global-mixing variants.

7. Limitations, scope, and recurrent misconceptions

Several limitations are explicit in the reported record. FUTransUNet is evaluated only on FUSeg; no external validation on additional DFU datasets or clinical environments is reported. The paper does not provide a numerical baseline comparison table against U-Net, vanilla TransUNet, DeepLab, SegNet, FUSegNet, or other segmentation architectures. Its empirical evidence is therefore a single-model report rather than a comparative benchmark study (Asare et al., 4 Aug 2025).

Reporting of implementation detail is also incomplete. The exact hardware configuration, software framework version, transformer embedding dimension, exact transposed-convolution setup, and explicit augmentation pipeline are not specified. Runtime, parameter count, FLOP count, and latency analysis are likewise absent. Failure cases and edge-case analyses are not shown (Asare et al., 4 Aug 2025).

Another recurring misconception is to interpret FUTransUNet as a universally established architecture name in the same sense as U-Net, TransUNet, or Swin-Unet. The current record is narrower. The name principally refers to the DFU segmentation model with Grad-CAM explainability (Asare et al., 4 Aug 2025). By contrast, search traffic for “FUTransUNet” may also surface unrelated or only loosely related models such as FocalUNETR or AFTer-UNet, which address different modalities and use distinct transformer-fusion mechanisms (Li et al., 2022, Yan et al., 2021).

Within its stated scope, FUTransUNet is best characterized as a hybrid CNN–ViT–U-Net for 2D ulcer segmentation, defined by local-global feature fusion and post hoc visual explanation rather than by volumetric attention, hierarchical transformer encoding, or explicit boundary-regression supervision.