SLSNet: Skin lesion segmentation using a lightweight generative adversarial network

1. Introduction

• Skin lesion segmentation is challenged by variation in lesion color, shape, texture, and size, along with fuzzy boundaries, blood vessels, hairs, and low contrast.Source: §1 · pp. 2–3
• Traditional methods can be inaccurate for fuzzy lesion boundaries and often depend on preprocessing such as hair removal and contrast enhancement.Source: §1 · p. 3
• MobileGAN targets accurate skin lesion segmentation with low computational cost, addressing the need for models suitable for resource-limited dermatoscopy devices.Existing methods may use hundreds of millions of parameters, while lightweight ENet produces lower accuracy than state-of-the-art methods.Source: §1 · pp. 3–4
• MobileGAN uses multiscale convolutional networks to extract lesion-relevant features and a 1-D kernel factorized network to reduce computational cost and resource use.Source: §1 · p. 4
• Position and channel attention mechanisms capture spatial and channel-feature correlations to improve discrimination between lesion and non-lesion representations.Source: §1 · p. 4
• The model combines multiscale aggregation for different lesion shapes with binary cross-entropy, Jaccard index, and L1-norm terms to address dermoscopic artifacts.Source: §1 · p. 4

Classical computer vision-based approaches

• Classical methods include thresholding, active contours, region growing, and clustering.Thresholding performs well mainly with apparent boundaries, while contour methods degrade with pigment changes, hair, and fuzzy lesion borders; clustering is inefficient for complex dermoscopic images.Source: §2 · p. 5

Deep learning-based approaches

• CNN segmentation advanced from FCN’s encoder–decoder design to U-Net’s skip connections and multiscale models such as FCRN.FCRN learns multiscale contextual features but remains vulnerable to low contrast, hair, and irregular lesion shapes.Source: §2 · pp. 5–6
• 79.87% IoU was achieved by a self-ensemble U-Net (Li et al., 2018) on ISBI 2017, compared with 72.8% for a noise-refined U-Net approach (Bissoto et al., 2018) on ISIC 2018.Other reported U-Net-related results include 76.7% IoU for SkinNet (Vesal et al., 2018) and 78.2% for SLSDeep on ISBI 2017.Source: §2 · p. 6
• GAN-based methods report 78.50% IoU for a ResNet-block model and 77.00% and 70.20% IoU for a two-stream model (Bisla et al., 2019) on ISBI 2017 and ISIC 2018.The two-stream approach jointly segments and classifies lesions and uses preprocessing to remove artifacts.Source: §2 · pp. 6–7

GAN-based approaches

• MobileNets uses depth-wise separable convolutions and has around 4 million parameters, while Mask-RCNN based on ResNet152 is described as a heavy model.Source: §2 · p. 7

Attention mechanism-based approaches

• Attention mechanisms weight informative channel or spatial features, suppress redundant responses, capture multiscale features at different positions, and can support faster learning with small datasets.These mechanisms have been applied in DCNNs,, and.Source: §2 · pp. 7–8

• Attention mechanisms contextualize local features and emphasize salient lesion structures while suppressing irrelevant image regions (Oktay et al., 2018).Source: §2 · p. 8
• Parameter-heavy skin-lesion segmentation models limit daily clinical use and real-time deployment on low-resource embedded systems.Source: §2 · p. 8
• MobileGAN is introduced as a lightweight, efficient GAN-based segmentation model for addressing the deployment limitations of heavier approaches.Source: §2 · p. 8
• Its factorized-attention module combines residual 1-D factorized-kernel convolution with channel attention to reduce computational parameters, while multiscale aggregation and position/channel attention capture lesions across scales and distinguish them from healthy skin.Source: §2 · pp. 8–9

3. Methodology

• The proposed architecture uses pix2pix GAN components, with an encoder–decoder generator mapping dermoscopic images to segmented lesions and a discriminator comparing generated and real masks.Figure 1 presents the generator and discriminator architecture.Source: §3 · p. 9, Figure 1 · p. 10
• A multiscale block aggregates coarse-to-fine dermoscopic features to reduce false detections caused by artifacts.The encoder generates input scales at 1/8, 1/4, and 1/2 resolution, upsamples lower-scale features, and averages them at the original size.Source: §3 · p. 9, §3.0.1 · pp. 9–10

3.0.1. The Encoder Network

• The encoder applies channel and position attention to model inter-channel dependencies and long-range spatial context for discriminative lesion representations.Position attention selectively aggregates contextual information across spatial locations, while channel attention highlights class-dependent feature maps.Source: §3.0.1 · pp. 10–11
• Residual 1-D kernel factorized layers replace standard 2-D filtering in factorized-attention modules to reduce computational complexity.The factorized modules feed the factorized features into channel attention, while the encoder progressively produces 64 and 128 feature maps before decoding the lesion boundaries.Source: §3.0.1 · pp. 10, 13
• The encoder includes a factorized-attention module with channel attention, and its architecture is presented as MobileGAN in Figure 4.Source: §3.0.1 · p. 14

3.0.2. The Decoder Network

• The decoder applies upsample-attention layers with PAM and FCM blocks to produce final feature maps, then upsamples them to obtain segmented images converted to binary masks using threshold 0.5.Source: §3.0.2 · p. 14

3.0.3. The Discriminator Network

• The discriminator uses four downsampling layers, adding PAM in the second and CAM in the third, then applies sigmoid to distinguish generated from real masks.The first three downsampling layers use 4×4 kernels, stride 2, and padding 1.Source: §3.0.3 · p. 14

3.1. Model training

• MobileGAN alternately trains the discriminator and generator using adversarial back-propagation with binary cross-entropy, L1, and Jaccard losses.The L1 term accelerates learning toward expected lesion boundaries, while the differentiable Jaccard loss promotes overlap with ground-truth masks and supports back-propagation.Source: §3.1 · pp. 14–15

Datasets

• The experiments used ISBI 2017 and ISIC 2018 separately, with ISBI split into 2,000 training, 150 validation, and 600 test images and ISIC using an 80/20 training-validation split.ISIC 2018 contains 2,594 images with expert-annotated masks; its validation and test sets contain 100 and 1,000 images without ground truth, respectively.Source: §4 · p. 16

Data augmentation

• Horizontal and vertical flips, gamma reconstruction, and CLAHE contrast changes expanded the training sets to 16,000 ISBI 2017 and 16,600 ISIC 2018 images.Source: §4 · p. 18

Implementation and image size

• Among the three evaluated image sizes—64 × 64, 128 × 128, and 256 × 256—128 × 128 produced the best segmentation results.Source: §4 · p. 18

Experimental results

• On the ISIC 2018 validation leaderboard, MobileGAN achieved the highest JSCth among the compared GAN-FCN and FrCN models, improving over them by 0.6% and 3.8%, respectively.Mask R-CNN and Rcnn-superpixels achieved slightly higher JSCth, but their parameter counts were not reported.Source: Table 2 · p. 19, §4 · pp. 19–21
• MobileGAN used 2.35 million parameters, compared with 10.61 million for GAN-FCN and 382.17 million for SegAN, making it substantially lighter than these models.The paper credits exploitation of 1-D kernels, PAM, and CAM with significantly reducing MobileGAN’s parameter count.Source: §4 · p. 20

Qualitative results

• About 95% accuracy was maintained on examples with similar lesion and skin colors, fuzzy boundaries, or tiny lesions, but segmentation became inaccurate when lesions occupied most of the image.The difficult cases lacked a proper boundary between lesion and healthy skin tissue, making precise delineation difficult for any segmentation model.Source: §4 · p. 21

Inference speed

• 8 ms (110 FPS) inference was achieved on a single GTX1080Ti GPU, with MobileGAN faster than GAN-FCN at every tested input scale.GAN-FCN required 6 ms more under the same configurations, supporting real-time operation on a single Mobile GPU.Source: Table 3 · p. 22, §4 · p. 22

Model variations

• The BL GAN baseline achieved 83.61% DSC and 72.93% JSC; adding PAM raised these to 86.01% and 75.96%, while adding CAM yielded 87.23% and 76.65%, respectively.Source: §4 · pp. 22–23
• 90.63% DSC and 81.98% JSC were achieved by MobileGAN with multiscale, PAM, and CAM blocks, outperforming all baseline GAN variations across five metrics.All models used 128×128 inputs, and the comparison is reported on the ISBI 2017 test dataset.Source: Table 4 · p. 23, §4 · p. 23

Input resolution

• At 64 × 64, the 8 × 8 final feature maps yielded very coarse information and lost important details, while 256 × 256 maps extracted lesion-irrelevant artifacts.These effects reduced segmentation accuracy and overall accuracy, respectively.Source: §4 · pp. 23–24

Loss-function analysis

• 74.48%, 76.80%, 79.88%, and 81.98% JSC resulted from progressively adding BCE, L1, and Jaccard losses, with the full combination significantly improving five metrics.L1 reduced sensitivity to outliers, while Jaccard loss detected subtle abnormalities missed by cross-entropy and reduced false positives.Source: Table 6 · p. 24, §4 · p. 24

• MobileGAN combines 1-D kernel factorized networks, multiscale aggregation, and position and channel attention within a GAN architecture.Source: §5 · p. 25
• MobileGAN was assessed on the ISBI 2017 test dataset and the ISIC 2018 validation dataset.Source: §5 · p. 25
• 78.4% threshold JSC was achieved on the ISIC 2018 validation dataset.Source: §5 · p. 25
• 2.35 million parameters gives MobileGAN a substantially reduced model size compared with state-of-the-art segmentation methods.Source: §5 · p. 25
• The authors propose future work involving a mobile application for segmenting lesions in images captured by low-resolution cameras.Source: §5 · pp. 25–26