Source-linked AI summary
DSNet: Automatic Dermoscopic Skin Lesion Segmentation
Md. Kamrul Hasan, Lavsen Dahal, Prasad N. Samarakoon, Fakrul Islam Tushar, Robert Marti Marly
TL;DR
Skin lesion segmentation is important for computer-aided melanoma diagnosis but remains challenging because lesion appearance and boundaries vary. The paper introduces DSNet, a lightweight semantic segmentation network, and reports state-of-the-art performance on ISBI and PH2 datasets while using fewer parameters than compared networks.
Problem
Skin lesion segmentation is a challenging but important component of computer-aided diagnosis because lesion appearance varies and segmentation supports downstream classification.
Method
The paper introduces DSNet, an automatic semantic segmentation network designed to reduce redundant encoder features and use a lightweight architecture.
Results
DSNet is reported to achieve state-of-the-art results on ISBI and PH2 while having the fewest parameters among compared networks.
Takeaways & Limitations
The proposed network is presented as a lightweight model that produces strong skin lesion segmentation results across the evaluated datasets.
Takeaways & Limitations
The authors identify further hyperparameter tuning and introducing more work as remaining directions.
Abstract
from arXiv · showhide
Automatic segmentation of skin lesion is considered a crucial step in Computer Aided Diagnosis (CAD) for melanoma diagnosis. Despite its significance, skin lesion segmentation remains a challenging task due to their diverse color, texture, and indistinguishable boundaries and forms an open problem. Through this study, we present a new and automatic semantic segmentation network for robust skin lesion segmentation named Dermoscopic Skin Network (DSNet). In order to reduce the number of parameters to make the network lightweight, we used depth-wise separable convolution in lieu of standard convolution to project the learned discriminating features onto the pixel space at different stages of the encoder. Additionally, we implemented U-Net and Fully Convolutional Network (FCN8s) to compare against the proposed DSNet. We evaluate our proposed model on two publicly available datasets, namely ISIC-2017 and PH2. The obtained mean Intersection over Union (mIoU) is 77.5 % and 87.0 % respectively for ISIC-2017 and PH2 datasets which outperformed the ISIC-2017 challenge winner by 1.0 % with respect to mIoU. Our proposed network also outperformed U-Net and FCN8s respectively by 3.6 % and 6.8 % with respect to mIoU on the ISIC-2017 dataset. Our network for skin lesion segmentation outperforms other methods and can provide better segmented masks on two different test datasets which can lead to better performance in melanoma detection. Our trained model along with the source code and predicted masks are made publicly available.
Conclusion
The paper concerns skin lesion segmentation, computer-aided diagnosis, melanoma detection, and deep learning.
- Skin lesion segmentation is identified as a central topic.
- Computer Aided Diagnosis (CAD) is connected to the paper’s application area.
- Melanoma detection is part of the clinical context.
1. Introduction
Skin lesion segmentation is important for computer-aided melanoma diagnosis but remains difficult because lesions vary substantially and may have indistinguishable boundaries and artifacts. The paper introduces DSNet, a lightweight semantic segmentation network intended to address these challenges and reports state-of-the-art results on ISIC and PH2 datasets.
- Motivation and problem: Skin lesion segmentation is a vital CAD component because classification features are obtained from the segmented region of interest.
- Motivation and problem: Automatic robust segmentation is challenging because lesions vary in skin color, texture, size, and boundary characteristics.
- Motivation and problem: Artifacts including hair, reflections, shadows, lighting variation, markers, and ink further complicate segmentation.
- Prior approaches: Traditional thresholding and region-based methods can fail with artifacts, noise, intensity variability, or fuzzy lesion borders.
- Prior approaches: Earlier learning-based approaches combine multi-stage networks, auto-context, full-resolution processing, dilated residuals, pyramid pooling, or dense encoder-decoder connections.
- Proposed contribution: DSNet is proposed as an automatic robust semantic segmentation network that reduces redundant encoder features and uses a lightweight design.
- Proposed contribution: The paper reports state-of-the-art results on the ISBI and PH2 datasets while describing DSNet as the most lightweight network producing such results.
2. Methods and Material
DSNet combines dense feature reuse, encoder–decoder processing, skip connections, and depth-wise separable convolutions for skin-lesion segmentation. It is evaluated on ISIC-2017 and PH2 datasets after addressing spatial-resolution and computational challenges.
- Architecture: DSNet uses an encoder–decoder architecture in which the encoder extracts discriminating features and the decoder projects lower-resolution features into pixel space.The decoder reconstructs a high-resolution predicted lesion mask.
- Design considerations: Sub-sampling improves spatial invariance but can cause resolution loss, coarse boundaries, checkerboard artifacts, and over-segmentation in predicted masks.The design uses skip connections and decoder processing to address these effects.
- Encoder: Dense connectivity gives each layer access to preceding feature maps, enabling feature reuse and making the encoder more compact.The encoder contains 121 layers and follows the DenseNet architecture.
- Decoder: Skip connections concatenate pooled encoder features with decoder feature maps to compensate for spatial information lost during sub-sampling.The connections are inspired by U-Net and use matching feature-map dimensions.
- Lightweight design: Depth-wise separable convolution replaces standard convolution in the decoder, reducing parameters by a factor of (1/NF + 1/K^2) per convolution layer.It performs channel-wise spatial convolution followed by a 1 × 1 point-wise convolution.
3. Results and Discussion
DSNet is evaluated against FCN8s, U-Net, and state-of-the-art networks using quantitative, qualitative, class-specific, and speed comparisons. It generally achieves stronger lesion segmentation with substantially fewer parameters and faster inference.
- DSNet improved mIoU over U-Net and FCN8s by 3.6% and 6.8%, respectively, when using the proposed loss function.
- DSNet used 3.8 and 13.8 times fewer parameters than U-Net and FCN8s, respectively.
- Qualitatively, FCN8s produced checkerboard artifacts, more false positives, and coarse boundaries, whereas U-Net showed under-segmentation and more false negatives.
- On PH2, which was not used for training or validation, DSNet achieved an mIoU of 0.87, supporting performance across datasets.
- DSNet achieved 0.595 seconds per image, compared with 9.7 seconds for FrCN, making it roughly 16 times faster in the reported comparison.
4. Conclusion
The conclusion presents DSNet and a combined binary cross-entropy–IoU loss for lightweight skin-lesion segmentation. Experiments support strong performance and efficiency, while future work targets tuning, augmentation, and broader medical applications.
- DSNet is introduced as a robust, automatic skin-lesion segmentation network evaluated through extensive experiments.
- The proposed loss combines binary cross-entropy with intersection over union and achieved higher true-positive rates in the reported experiments.
- The training challenge of limited labeled lesion images was addressed by transferring knowledge from a pretrained model and augmenting images.
- DSNet has the fewest parameters among the compared networks while outperforming baseline segmentation networks on several metrics.
- Further hyperparameter tuning and additional color-space augmentation may improve segmentation performance, and generality in other medical contexts remains to be verified.
Appendix A. Prediction Samples
Prediction samples overlay predicted masks with ground truth and distinguish true positives, false positives, and false negatives. The samples include challenging images and the network’s worst predictions.
- The samples overlay predicted segmentation masks on ground truth, with true positives, false positives, and false negatives shown in green, yellow, and red.
- The first six rows show challenging images, while the last three rows show the worst predictions obtained using DSNet.