DTONet a Lightweight Model for Melanoma Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning

Background:

• Melanoma segmentation accuracy improves with complex neural networks and abundant hardware.
• Limited computational resources in clinical settings and for the public hinder the application of high-complexity models.
• There is a need for efficient deep learning models for melanoma segmentation that perform well on resource-constrained devices.

Purpose of the Study:

• To propose a lightweight deep learning network for accurate melanoma segmentation with minimal resource consumption.
• To introduce DTONet (double-tailed octave network) as an efficient solution for melanoma segmentation.
• To demonstrate the superior performance and generalization capability of DTONet compared to existing models.

Main Methods:

• Development of DTONet, a novel lightweight deep learning architecture.
• Comparative analysis of DTONet's computational parameters against mainstream models like UNet.
• Validation of DTONet's segmentation accuracy and generalization on the PH2 dataset.

Main Results:

• DTONet possesses a significantly reduced parameter count (30,859), approximately 1/256th of UNet.
• The proposed network achieves superior segmentation accuracy, indicated by improved IOU scores over similar models.
• DTONet demonstrated excellent generalization capabilities, outperforming existing models on the PH2 dataset.

Conclusions:

• DTONet offers a highly efficient and accurate solution for melanoma segmentation.
• The lightweight design makes DTONet suitable for deployment in resource-limited environments.
• The model shows strong potential for practical applications in clinical settings and public health initiatives.