Dermoscopic Image Segmentation via Multistage Fully Convolutional Networks.

Area of Science:

• Medical image analysis
• Computer-aided diagnosis
• Dermatology

Background:

• Automated skin lesion segmentation is crucial for melanoma diagnosis.
• Existing methods struggle with lesions exhibiting fuzzy boundaries, low contrast, or artifacts.
• Current techniques often require extensive parameter tuning and preprocessing.

Purpose of the Study:

• To develop an automated skin lesion segmentation method using fully convolutional networks (FCNs).
• To address limitations of standard FCNs in segmenting challenging lesions with fuzzy boundaries or low texture contrast.
• To improve the accuracy and boundary definition of segmented skin lesions.

Main Methods:

• Leveraged multistage fully convolutional networks (FCNs) to learn complementary visual characteristics.
• Employed early-stage FCNs for coarse appearance and localization, and late-stage FCNs for subtle boundary details.
• Introduced a parallel integration method to combine outputs from different FCN stages for refined segmentation.

Main Results:

• Achieved an average Dice coefficient of 91.18% on the ISBI 2016 Skin Lesion Challenge dataset.
• Obtained an average Dice coefficient of 90.66% on the PH2 dataset.
• Demonstrated superior performance compared to state-of-the-art methods on public benchmark datasets.

Conclusions:

• The proposed multistage FCN approach significantly enhances skin lesion segmentation accuracy.
• The method effectively handles challenging lesions, providing accurate localization and well-defined boundaries.
• This technique offers a more robust and effective solution for automated skin lesion analysis.