[Paper Review] SLSDeep: Skin Lesion Segmentation Based on Dilated Residual and Pyramid Pooling Networks

Md. Mostafa Kamal Sarker, Hatem A. Rashwan+9 more|arXiv (Cornell University)|May 25, 2018

Cutaneous Melanoma Detection and ManagementMedicine22 references114 citations

TL;DR

SLSDeep proposes an encoder-decoder model using dilated residual networks and pyramid pooling for skin lesion segmentation, optimized with a combined NLL+EPE loss, achieving state-of-the-art results on ISBI 2016/2017 datasets.

ABSTRACT

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.

Motivation & Objective

• Motivate automated melanoma segmentation to aid diagnosis and address variability in dermoscopic images such as size, shape, and boundary fuzziness.
• Develop an encoder-decoder architecture that captures coarse-to-fine features without heavy pre/post-processing.
• Investigate a loss function that combines Negative Log Likelihood (NLL) with End Point Error (EPE) to improve boundary sharpness.

Proposed method

• Encoder based on dilated residual networks to extract hierarchical features.
• Decoder employing a pyramid pooling layer to generate multi-scale context before reconstruction.
• Concatenation of encoder features with decoder pathways via a targeted skip-connection strategy (best with a single skip from encoder end to PPN).
• A loss function that combines NLL (for binary segmentation probability) with EPE (for boundary preservation): L_total = L_log + α L_epe with α = 0.5.
• Training uses PyTorch, Adam optimizer, 0.001/0.01 learning rates for encoder/decoder, data augmentation (scaling and rotation), batch size 16, 100 epochs.
• Evaluation on ISBI 2016 and 2017 dermoscopy datasets using ACC, DIC, JAC, SEN, SPE.

Experimental results

Research questions

• RQ1 Can a dilated residual encoder combined with pyramid pooling in the decoder improve SLS accuracy and boundary sharpness over prior methods?
• RQ2 Does a loss function that combines Negative Log Likelihood and End Point Error outperform standard cross-entropy for binary skin lesion segmentation?
• RQ3 What is the impact of skip-connection strategies on segmentation performance in this architecture?

Key findings

• SLSDeep achieves top performance on ISBI 2016 with ACC 0.984, DIC 0.955, JAC 0.913, SEN 0.945, SPE 0.992.
• On ISBI 2017, SLSDeep attains ACC 0.936, DIC 0.878, JAC 0.782, SEN 0.816, SPE 0.983.
• Compared to variants, SLSDeep with NLL+EPE and a single encoder–PPN skip connection consistently improves DIC and JAC (e.g., ISBI 2016: +3% DIC, +5% JAC relative to variants).
• SLSDeep-EPE and SLSDeep+ASC show mixed results, with EPE generally improving boundary preservation but not always overall metrics.
• SLSDeep runs at over 100 images per second on a modern GPU for 384x384 inputs, demonstrating practicality for large-scale analysis.