EM-Net: Effective and Morphology-aware Network for Skin Lesion Segmentation

Abstract

Dermoscopic images are essential for diagnosing various skin diseases, as they enable physicians to observe subepidermal structures, dermal papillae, and deeper tissues otherwise invisible to the naked eye. However, segmenting lesions in these images is challenging due to their irregular boundaries and significant variability in lesion characteristics. To address these challenges, we propose a effective and morphology-aware network that utilizes a hybrid feature extractor combining CNN and ViT architectures. At the same time, we enhanced the proposed segmentation model. Specifically, we propose a boundary delineation component that uses a non-convex optimization function for learning general representations and accurately delineates lesion boundaries, thus enhancing the extraction of details. Additionally, we also introduce an adaptive segmentation strategy through the integration of the few-shot domain generalization module to improve the model's generalization across different datasets. Validation on multiple publicly available dermoscopic image datasets, including ISIC, PH², PAD-UFES-20, and the University of Waterloo skin cancer database, demonstrates that our method achieves state-of-the-art performance with significant improvements in Dice, Acc, Pre, IoU, and Re. These results confirm the robustness and adaptability of our model. The code is available at: https://github.com/Bean-Young/EM-Net.

Highlights

Develop a Morphology-aware Module for lesion boundary capture.
Introduce Few-shot Domain Generalization for robust segmentation.
Build a CNN-ViT hybrid framework for skin lesion segmentation.

Method

Overall architecture of the proposed EM-Net.

Morphology-aware Module

Illustration of morphology-aware boundary capture.

Results

Quantitative Results

Method	test-ISIC 2016					test-ISIC 2017					validation-ISIC 2018
Method	Dice	IoU	Acc	Pre	Re	Dice	IoU	Acc	Pre	Re	Dice	IoU	Acc	Pre	Re
U-Net	90.29	82.58	94.76	91.29	87.72	82.80	73.30	92.30	89.07	79.30	88.43	79.71	93.16	89.46	85.89
Att-UNet	90.83	83.43	94.43	92.12	88.42	83.20	74.40	92.50	87.29	84.11	88.32	79.52	93.17	87.73	87.11
CE-Net	91.80	85.01	95.30	91.84	90.32	85.61	77.54	93.51	89.91	84.52	89.45	81.32	94.03	89.18	87.60
CPF-Net	91.34	84.24	95.18	91.34	89.78	84.70	76.20	93.00	86.67	83.33	88.86	80.45	94.68	88.44	87.57
MS RED	91.43	84.43	95.67	91.60	89.53	84.83	76.32	93.10	87.07	83.74	89.18	80.92	95.03	89.14	87.36
FAT-Net	91.49	84.49	96.07	91.11	90.41	85.01	76.92	93.64	87.77	84.52	89.18	80.92	95.18	88.52	87.88
TransUNet	91.32	84.89	96.21	92.28	89.63	85.51	77.34	93.47	88.56	84.22	89.50	82.61	93.67	88.12	89.50
Swin-UNet	91.27	84.14	96.18	91.39	89.45	83.50	72.28	93.20	84.27	83.14	88.46	79.86	94.45	88.24	87.43
Ours	91.89	85.57	96.25	95.65	89.66	86.42	78.37	93.97	92.17	85.33	91.47	84.78	95.35	91.28	93.15

Comparison with alternative approaches on ISIC datasets.

Method	test-ISIC 2018
Method	Dice	IoU	Acc	Se	Sp
SCDC	88.50	82.40	92.90	95.30	91.10
ACA-Net	89.10	81.90	-	94.30	93.20
Deeplabv3+	89.60	82.50	94.20	96.20	92.10
SESV-DLab	90.20	83.30	94.60	96.20	92.50
ICL-Net	90.30	83.90	94.40	94.10	92.90
Ours	90.30	83.60	94.70	92.40	93.90

Comparison on the ISIC 2018 Challenge ranking list.

Method	test-PH²
Method	Dice	IoU	Acc	Pre	Re
U-Net	90.56	83.56	94.86	90.30	93.47
Att-UNet	90.29	82.70	94.68	91.35	93.26
CE-Net	90.87	83.90	95.38	93.21	96.01
CPF-Net	91.67	85.48	95.59	91.81	95.90
MS RED	92.65	85.29	95.46	90.38	95.52
FAT-Net	92.21	85.18	95.43	92.85	96.33
TransUNet	90.96	83.99	95.42	91.18	95.68
Swin-UNet	92.69	87.08	96.03	91.42	94.87
ICL-Net	92.80	87.25	96.32	-	95.46
Ours	94.03	88.92	96.34	94.06	94.57

Quantitative evaluation on the PH² dataset.

Method	test-Waterloo					test-PAD
Method	Dice	IoU	Acc	Pre	Re	Dice	IoU	Acc	Pre	Re
U-Net	82.71	72.41	96.03	74.78	96.47	87.71	79.45	94.85	82.47	95.69
TransUNet	86.51	77.21	97.68	80.12	96.49	90.82	83.80	96.98	90.76	92.06
Ours	90.03	82.47	98.42	94.69	87.14	91.88	85.37	97.41	95.69	89.02

Comparison on datasets imaged by mobile phones.

Qualitative Results

Visual comparison on the ISIC dataset.

ISIC 2016 and PH2 segmentation comparison

Visual comparison on ISIC 2016 and PH² datasets.

Mobile phone dataset segmentation comparison

Visual comparison on mobile-phone dermoscopic images.

Citation

@article{zhu2025net,
  title={EM-Net: Effective and morphology-aware network for skin lesion segmentation},
  author={Zhu, Kaiwen and Yang, Yuezhe and Chen, Yonglin and Feng, Ruixi and Chen, Dongping and Fan, Bingzhi and Liu, Nan and Li, Ying and Wang, Xuewen},
  journal={Expert Systems with Applications},
  pages={127668},
  year={2025},
  publisher={Elsevier}
}