Convolutional neural networks (CNNs) remain the foundation of most methods used to segment skin lesions. Classic architectures such as U-Net [5] and its derivatives have been widely extended to improve pixel-level accuracy. For example, UNet++ [6] and DFF-UNet [7] integrated deep feature fusion, whereas MRP-UNet [17] incorporated multiscale input fusion and pyramid dilated convolutions to capture lesions of varying sizes better. These fully convolutional networks (FCNs) highlight the strength of CNNs in local feature learning; however, they often struggle to effectively capture the global context. To address this, several models have begun to embed attention- or transformer-based modules in CNN frameworks.

Transformer-based methods address the limitations of CNN by capturing long-range dependencies and local features. Mas-TransUNet [18] incorporates CNN encoding and transformer modules for simultaneous global and local feature extractions. DuaSkinSeg is a two-encoder-based method using MobileNetV2 and Vision Transformer to utilize convolutional features with contextual representation [19]. BDFormer employs a boundary-aware dual-decoder transformer to improve edge accuracy [20], and EM-Net incorporates a morphology-aware design to better represent the lesion structure [21]. These approaches enhance segmentation performance, but often introduce greater model complexity, limiting their suitability for real-time or resource-constrained environments. However, their computational demands hinder their automatic clinical use in real-time applications, which require speed and hardware efficiency.

One way to improve the precision and performance is to introduce an attention mechanism. Attention Squeeze U-Net focuses on embedded devices [22]. This is not due to the dataset size, but to the deployment environment, as large datasets are mostly required only for offline model training. Attention networks are intrinsically lightweight, which improves their performance, while making them more compact. While alternative efficient methods, such as boundary-aware attention and a hybrid CNN-transformer encoder-decoder, enhance the fusion of region features, they also introduce additional computational overhead [11]. They identified a trade-off between segmentation accuracy and clinical applicability.

Lightweight versions are popular for real-time applications. DSNet is a lightweight CNN with depthwise convolutions and boundary-aware loss, achieving strong segmentation performance with fewer parameters [23]. In addition to architectural design, a model explores compression techniques such as pruning, quantization, and knowledge distillation to reduce inference time and memory usage. As such, “lightweight” in clinical settings involves a smaller parameter count and faster runtime on mobile or embedded systems, where GPUs are not available. Several strategies based on swine transformers with parameter sharing and anti-aliasing upsampling have been investigated; however, challenges remain in the ambiguous segmentation of lesions [24]. The use of lightweight designs allows point-of-care implementation to be integrated into research models.

This categorization emphasizes incremental developments in CNN-based, transformer-based, and lightweight methods, thereby providing high segmentation accuracy and clinical utility trade-offs. Despite notable progress in lesion segmentation, existing approaches face three significant challenges: (i) high computational costs that limit real-time use, (ii) reduced accuracy for small- or low-contrast lesions, and (iii) limited adaptability in resource-constrained clinical environments. These limitations indicate the importance of a lightweight yet accurate model. In the following section, LANET explain that lightweight designs help bridge the gap between research models and point-of-care deployment. These gaps were addressed by combining depthwise convolutions, a customized attention block, and an optimized ASPP module.

Table 1 lists the human-to-machine ratio (HAM10000), which consists of images and the ground truth [25]. The dataset consists of 10,015 images organized into seven categories, with a resolution of 600 × 450 pixels in the JPEG format. The ground-truth dataset is the same size, but it is in binary images in the PNG format. Table 2 lists other public datasets used in this study. The ISIC 2017 and 2018 datasets are widely used as benchmarks for skin cancer detection and consist of dermoscopic images. The ISIC 2017 comprises 2750 images and supports three tasks: lesion segmentation, dermoscopic feature identification, and lesion categorization into melanoma, seborrheic keratosis, and benign nevi [16]. The ISIC 2018 dataset comprises 3694 images. It is used for multiclass classification across seven diagnostic categories [26]. Both datasets offer expert annotations and have established themselves as benchmarks for assessing DL models in automated skin-lesion analyses. The PH2 dataset consists of 200 skin lesion images, each with its corresponding label, at a resolution of 768 × 560 [27]. It is commonly used to assess the generality and efficacy of a model. Fig. 1 shows the sample images and corresponding ground-truth masks from all four datasets.

LANET is based on U-Net [5]. The encoder-decoder architecture was designed for efficiency and segmentation accuracy (Fig. 2). It includes three main components: (i) an encoder that incorporates attention mechanisms and grouped convolutions; (ii) a decoder; and (iii) an ASPP module for multiscale contextual understanding. The model accepts three channels (R, G, and B) as input for an image of size 224 × 224 × 3, and outputs a pixel-wise segmentation map with two classes (lesion and background).

To ensure the robustness and reliability of the proposed model, the LANET was evaluated on the HAM10000 dataset using 5-fold cross-validation. Data were randomly subdivided into five subsets. In each cycle, four subsets were trained in, and one was validated such that each image was validated exactly once. Hence, the final performance measures reported the average and standard deviation across all folds, providing a better estimate of how well the model can be generalized in practice.

To better illustrate the learning behavior of the model, we tracked the accuracy and loss of both the training and validation sets across all the iterations. As shown in Fig. 7, the curves demonstrate stable learning, with consistent trends between the two sets. The validation accuracy closely followed the training accuracy, and the loss decreased smoothly for both, indicating an effective convergence without overfitting. The final validation accuracy reached 96.12%, which was consistent with the quantitative results reported in the main experiments. This comparison confirms that LANET generalizes well to unseen data and maintains reliable performance throughout the training process.

Table 4 presents the comparison criteria for testing these models using the LANET. All GFLOPs are computed for an input of 224 × 224 using the convention that one multiply–accumulate (MAC) equals two floating-point operations. Compared to U-Net (7.42 GFLOPs) and U-Net++ (8.45 GFLOPs), the proposed LANET achieves higher segmentation accuracy (as mentioned in the results section) while requiring only 4.22 GFLOPs. This highlights superior computational efficiency of LANET.(4)Sizeofmemory(inMegaBytes)=Nparametres×4bytes10242

Sizeofmemory=846786×410242=33871441048576=3.2 MB. Assuming 32-bit float precision (4 bytes per parameter), a LANET with 846,786 parameters requires approximately 3.2 MB of memory.

Five primary metrics were used to evaluate the performance of the LANET. Evaluation metrics are beneficial for the overall analysis and offer helpful information regarding many characteristics of the segmentation results. Eqs. (5)–(9) show the subsequent metrics. Where: “TP: True Positive, TN: True Negative, FP: False Positive, FN: False Negative” (5)Sensitivity=TPTP+FN (6)Specificity=TNTN+FP (7)Accuracy=TP+TNTP+TN+FP+FN (8)DiceCoefficient(Dice)=2TP2TP+FP+FN (9)IntersectionoverUnion(IoU)=TPTP+FP+FN

Table 5 evaluation of our method on the HAM10000 dataset LANET achieved accuracies of 96.44%, Dice = 94.53% and IoU = 92.61% over the HAM10000 dataset (shown in Table 5). All these results are higher than U-Net, U-Net++, DeepLabv3+, and MRP-UNet. The model shows a 17.3-point gain in Dice over U-Net and a 5.4-point gain over U-Net++. Compared with MRP-UNet, LANET improves Dice by 1.58 points, confirming its stronger boundary detection ability. A five-fold cross validation was conducted. LANET achieved an average Dice of 96.4% ± 0.6% and IoU of 89.1% ± 0.7%, showing stable performance. The results in Fig. 8 compare the five SOTA metrics. Fig. A1 in Appendix B shows the predictions in green color and the ground truth in red color.

The LANET returned a Dice score of 94.53% and accuracy of 96.44%, revealing competitive in-domain performance. The results transferred to ISIC 2017, ISIC 2018, and PH2 for LANET were the best with Dice scores from 90.61% to 91.09% and accuracies over the threshold, and all datasets were higher than 93%, showing a stable performance in segmentation over polling positions including different datasets.

LANET obtained the highest accuracy (96.78%) and one of the best Dice scores (92.90%) among all models. Its Dice score exceeds those of U-Net, U-Net++, and U-Net+3 by 6.7, 5.2, and 5.6 points, respectively. The sensitivity (90.30%) and specificity (92.83%) showed that the model accurately detected lesion areas, while limiting false positives. The results were statistically significant (p < 0.05). The summary statistics of the performance of the models are shown in Fig. 9. Qualitative examples demonstrate the superior performance of LANET in tracing lesion borders compared to other competitors. Fig. A2 in Appendix B shows some segmented image samples. Qualitative examples support the quantitative findings, showing accurate segmentation across various lesion shapes and sizes.

Table 7 presents the comparative results of various DL models against the LANET. LANET achieved the highest overall accuracy (96.32%) and Dice score (92.21%) for the ISIC 2018. Its Dice score surpasses U-Net by 17.1 points and exceeds DSU-Net and EM-Net by roughly 2 points. The sensitivity (90.84%) and specificity (98.22%) demonstrated a balanced discrimination between the lesion and background regions. A statistical comparison of the performances is shown in Fig. 10. Across all SOTA models, LANET produced the second-highest specificity, while maintaining competitive sensitivity. Examples of the representative segmentations are shown in Fig. A3 in Appendix B. The qualitative outputs showed an improved boundary precision and reduced noise sensitivity.

Table 8 lists the performance of the PH2 dataset. LANET achieved the highest accuracy (97.89%), followed by strong Dice (93.70%) and IoU (91.60%). It improved Dice over U-Net by 18.4 points, U-Net++ by 5.3 points, and DeepLabv3+ by 2.7 points. The LANET also achieved a high sensitivity (95.80%) and specificity (92.90%), indicating reliable lesion recognition with few false alarms. The model consistently outperformed the classical and transformer-based methods (p < 0.05). The qualitative samples showed accurate structural preservation even in challenging low-contrast lesions. The qualitative examples in Fig. A4 in Appendix B shows that the LANET is more accurate at boundary rendering and lesion localization than the other models.

A cross-dataset experiment was performed using LANET to evaluate its generalizability to other imaging datasets, as shown in Table 9. The model was trained on the dataset HAM10000 (refer to Table 5) and tested on three non-introduced datasets (ISIC 2017, ISIC 2018, and PH2) without any retraining or fine-tuning. These datasets are diverse with respect to illumination, lesion type, and color normalization. Cross-dataset testing shows that the LANET maintains a strong performance when applied to datasets with different imaging characteristics, indicating good generalization.

The LANET model uses interpretability tools (i.e., convolutional activation maps and Grad-CAM) to illustrate the image regions that drive its predictions. These visualizations are concentrated in the lesion regions rather than irrelevant artifacts to mitigate clinical trust hesitation and provide transparency. Because DL models are considered “black boxes,” these interpretability techniques shed light on what LANET has thought of. Through a performance assessment using the Dice coefficient, we demonstrated that the LANET ensures a good overlap between the predictions and ground-truth masks for multiple datasets. From Table 10, the high Dice scores show reliable and accurate segmentation and generalization, even when considering the imbalance between classes and slight boundary changes.