DFU-FuzzyLiteUNet: A lightweight u-net with Fuzzy sigmoid and lite transformer for diabetic foot ulcer segmentation

DFU-FuzzyLiteUNet is a lightweight image segmentation model proposed for Diabetic Foot Ulcer (DFU) analysis to support critical needs in healthcare and accurate early detection. To improve the segmentation accuracy , a fuzzy approach is applied through a Fuzzy Sigmoid activation function in the encoder part, which enables a smoother and more significant feature capture. In addition, Smooth Transformer Block is used to efficiently capture the spatial relationship between image areas. The data processed by the transformer is passed to the Decoder Block through an Attention Gate at the skip connection, which ensures only relevant features are passed to improve the accuracy of segmentation prediction. The model was evaluated on two public datasets, namely FUSC 2021 and DFUC 2022, as well as one primary dataset that has been annotated by medical professionals. The evaluation results showed superior performance, with a Dice Coefficient of 91.6 % on the FUSC 2021 test data, an improvement of 3.83 % over the previous best model. The ablation study confirmed that the combination of Fuzzy Sigmoid and ReLU activation functions surpassed other activation functions. The model remains efficient with only 2.16 G FLOPS, a size of 2.35 MB, and 615 K parameters, resulting in a 55.92 % G FLOPS reduction over previous studies. Feature map visualization further validated the model’s ability to extract relevant segmentation features, showing great potential for clinical applications.