FAC-Net: Feedback Attention Network Based on Context Encoder Network for Skin Lesion Segmentation.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Skin lesions are early indicators of skin cancer, making their accurate segmentation crucial.
• Current data augmentation techniques in dermatological datasets create excessive parameters, hindering practical AI applications in medicine.

Purpose of the Study:

• To propose an effective feedback attention network (FAC-Net) for precise skin lesion segmentation.
• To address the limitations of data augmentation in dermatological AI by developing a parameter-efficient model.

Main Methods:

• Developed FAC-Net, incorporating a feedback fusion block (FFB) and an attention mechanism block (AMB).
• Evaluated the model on public datasets (ISIC2018, ISBI2017, ISBI2016) using Jaccard index (JA) and Dice coefficient (DC) metrics.
• Focused on achieving high segmentation accuracy without relying on data enhancement.

Main Results:

• FAC-Net achieved a Dice coefficient (DC) of 91.19% and a Jaccard index (JA) of 83.99% on the ISIC2018 dataset.
• Demonstrated over 1% improvement in DC and JA compared to baseline networks on ISIC2018.
• Showcased improved segmentation metrics across all tested public datasets.

Conclusions:

• FAC-Net offers superior skin lesion segmentation performance compared to many deep learning architectures.
• The lightweight and efficient design of FAC-Net enables practical application in smart assisted medicine.
• The proposed model effectively extracts rich feature mappings without data augmentation, setting a new standard for dermatological image analysis.