Diabetic foot ulcers are chronic complications related to infection, amputation, and increased mortality. Automatic segmentation of wound areas is essential for evidence-based monitoring, but is challenging due to blurred lesion boundaries, variations in lighting and texture, and computational limitations in clinical devices. This study introduces MFLiT-UNet, a lightweight segmentation architecture that combines FuzzyEncoderBlock on the encoder path, SmoothingTransformerBlock as a Transformer bottleneck, and a depthwise separable decoder with attention gate. To overcome class imbalances as well as edge obscurity, we use Hybrid Loss which combines Tversky Loss and Fuzzy Dice. Evaluations on FUSC 2021, DFUC 2022, and InWCCA primary datasets show high performance in source domains and remain competitive across domains. At FUSC 2021, MFLiT-UNet achieved DC 92.65 % and IoU 86.31 %, surpassing Mobile U-Net by +4.92 DC points and +8.11 IoU points. In addition, our model outperformed DFU-MambaLite in the same test scenario, demonstrating better segmentation quality (higher DC/IoU) in the corpus. The ablation study confirmed significant efficiency: compared to the basic Conv2D decoder, the lightest variant lowered the number of parameters by 66.35 %, the model size 65.46 %, and the FLOPs from 3.71 G to 1.83 G, with an average inference time of about 20 to 30 milliseconds per image. These findings suggest that the integration of fuzzy elements at the feature and objective levels, coupled with the global context of the economical Transformer, results in a good balance between accuracy and efficiency. MFLiT-UNet is eligible for consideration for application on limited-resource devices to support clinical decision-making in DFU cases.