Source-linked AI summary
SLSDeep: Skin Lesion Segmentation Based on Dilated Residual and Pyramid Pooling Networks
Md. Mostafa Kamal Sarker, Hatem A. Rashwan, Farhan Akram, Syeda Furruka Banu, Adel Saleh, Vivek Kumar Singh, Forhad U H Chowdhury, Saddam Abdulwahab, Santiago Romani, Petia Radeva, Domenec Puig
TL;DR
Skin lesion segmentation remains challenging because dermoscopic images have variable shapes, colors, fuzzy boundaries, and artifacts that limit conventional methods. SLSDeep addresses this with a dilated-residual encoder, pyramid-pooling decoder, and combined NLL-EPE loss, outperforming evaluated methods on two ISBI datasets.
Problem
Dermoscopic lesion segmentation is difficult because lesions vary in size, shape, color, and boundary clarity, while conventional methods can be unreliable on inhomogeneous images and depend on preprocessing.
Method
SLSDeep uses an encoder-decoder architecture combining dilated residual networks, pyramid pooling, skip connections, and an NLL-EPE loss for skin lesion segmentation.
Results
SLSDeep outperformed evaluated methods on the ISBI 2016 and 2017 datasets across multiple segmentation metrics, including accuracy, Dice, Jaccard, and specificity.
Takeaways & Limitations
The model provides a robust skin lesion segmentation technique across two public ISBI benchmarks using multiple evaluation metrics.
Takeaways & Limitations
Evaluation is limited to the ISBI 2016 and 2017 datasets, with testing on other color spaces and medical applications left for future work.
Abstract
from arXiv · showhide
Skin lesion segmentation (SLS) in dermoscopic images is a crucial task for automated diagnosis of melanoma. In this paper, we present a robust deep learning SLS model, so-called SLSDeep, which is represented as an encoder-decoder network. The encoder network is constructed by dilated residual layers, in turn, a pyramid pooling network followed by three convolution layers is used for the decoder. Unlike the traditional methods employing a cross-entropy loss, we investigated a loss function by combining both Negative Log Likelihood (NLL) and End Point Error (EPE) to accurately segment the melanoma regions with sharp boundaries. The robustness of the proposed model was evaluated on two public databases: ISBI 2016 and 2017 for skin lesion analysis towards melanoma detection challenge. The proposed model outperforms the state-of-the-art methods in terms of segmentation accuracy. Moreover, it is capable to segment more than $100$ images of size 384x384 per second on a recent GPU.
1 Introduction
Skin lesion segmentation remains difficult because dermoscopic images vary in appearance and lesion boundaries can be fuzzy, while traditional methods depend on fragile preprocessing. SLSDeep addresses these challenges with an encoder-decoder model combining dilated residual networks, pyramid pooling, skip-connections, and a loss based on NLL and EPE.
- Dermoscopic images exhibit varying sizes, shapes, colors, fuzzy boundaries, and hair, making melanoma segmentation challenging.
- Traditional thresholding, edge-based, region-based, clustering, and supervised methods are unreliable for inhomogeneous images or blurred lesion boundaries.
- These traditional methods also depend on preprocessing such as filtering, illumination correction, and hair removal, which can reduce generalizability.
- SLSDeep combines skip-connections, dilated residual networks, and pyramid pooling in an encoder-decoder architecture for skin lesion segmentation.
- Its loss combines Negative Log Likelihood and End Point Error to detect melanoma regions while preserving sharp edges.
2 Proposed Model
The proposed model uses dilated residual layers in its encoder and pyramid pooling in its decoder to combine coarse-to-fine image features. Its loss combines NLL with EPE so segmentation training accounts for both classification and boundary preservation.
- 2.1 Network Architecture: The encoder uses an initial convolution and pooling stage followed by four pretrained dilated residual blocks that extract progressively larger feature maps.The encoder output is 1/8 of the input resolution, producing 48 × 48 feature maps from a 384 × 384 image.
- 2.1 Network Architecture: The decoder uses pyramid pooling at 1 × 1, 2 × 2, 3 × 3, and 6 × 6 scales to produce and concatenate multi-scale feature maps.
- 2.2 Loss Function: Cross-entropy can be biased toward the background when melanoma occupies a small image region, motivating a combined NLL and EPE loss.
- 2.2 Loss Function: NLL serves as the objective loss for binary classification using the true label and estimated positive-class probability.
- 2.2 Loss Function: EPE compares the generated mask with ground truth through derivative-based absolute error to preserve object boundaries.
- 2.2 Loss Function: The total loss is Ltotal = Llog + αLepe, with α < 1 and α = 0.5 in this work.
3 Experimental Setup and Evaluation
SLSDeep was evaluated on two public ISBI datasets using standard segmentation metrics, ablations, and qualitative examples. It outperformed evaluated methods on most reported measures, while qualitative results showed accurate boundaries alongside errors on difficult cases.
- Experimental setup: The model was evaluated on ISBI 2016 and ISBI 2017 dermoscopy datasets using specificity, sensitivity, Jaccard, Dice, and accuracy.ISBI 2016 contains 900 training and 379 testing images; ISBI 2017 contains 2000 training, 150 validation, and 600 testing images.
- Experimental setup: The evaluation compared SLSDeep with variants that removed EPE or used all encoder-decoder skip connections.The final model used NLL and EPE with one skip connection between the encoder’s last layer and the PPN layer.
- Quantitative results: On ISBI 2016, SLSDeep achieved the best results among evaluated methods across five metrics, with around 3% higher DIC, 8% higher JAC, and over 98% overall accuracy.The comparison included ExB and three other top methods identified in the paper.
- Quantitative results: On ISBI 2017, SLSDeep improved DIC by 3% and JAC by 2% while outperforming three benchmark methods on ACC, DIC, and JAC.The benchmark method [25] achieved the highest SEN, exceeding SLSDeep by 0.9%.
- Ablation results: Using EPE improved DIC and JAC by 3% and 5% on ISBI 2016 and by 5% and 8% on ISBI 2017.SLSDeep with one skip connection also outperformed the all-skip-connection variant on both datasets.
- Qualitative results: Qualitative examples showed accurate segmentation with sharp edges, but also false regions and larger false-negative areas for difficult lesions.The most difficult example involved a very small lesion region.
4 Conclusions
The paper concludes that SLSDeep is a robust encoder-decoder model for skin lesion segmentation across two ISBI datasets and several evaluation metrics. Future work will examine different color spaces and applications to other medical tasks.
- Conclusion: SLSDeep combines dilated residual encoding with pyramid pooling and upsampling to reconstruct segmented dermoscopic images.The paper describes this as an encoder-decoder approach for skin lesion segmentation.
- Conclusion: SLSDeep outperformed the evaluated literature methods on the ISBI 2016 and ISBI 2017 datasets across reported segmentation measures.The conclusion specifically names accuracy, Dice coefficient, Jaccard index, and specificity.
- Future work: Future work will explore different color spaces and apply the model to other medical applications to assess its versatility.