EffiSegNet: Gastrointestinal Polyp Segmentation through a Pre-Trained EfficientNet-based Network with a Simplified...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Image segmentation is crucial for medical diagnosis.
• Traditional U-shaped networks can be computationally intensive.
• Transfer learning offers a promising approach to enhance model performance.

Purpose of the Study:

• To introduce EffiSegNet, an efficient segmentation framework.
• To leverage transfer learning with pre-trained Convolutional Neural Network (CNN) backbones.
• To improve gastrointestinal polyp segmentation accuracy and reduce computational cost.

Main Methods:

• Developed EffiSegNet with a simplified decoder and full-scale feature fusion.
• Employed transfer learning using pre-trained CNN classifiers as backbones.
• Evaluated the model on the Kvasir-SEG dataset for gastrointestinal polyp segmentation.

Main Results:

• EffiSegNet-B4 achieved state-of-the-art performance with an F1 score of 0.9552 and mIoU of 0.9056.
• Achieved the highest reported scores on the Kvasir-SEG dataset using a pre-trained backbone.
• Training from scratch also yielded competitive results, outperforming previous work.

Conclusions:

• EffiSegNet demonstrates the effectiveness of transfer learning in image segmentation.
• A well-designed encoder is critical for high-performing segmentation networks.
• The proposed framework offers an efficient and accurate solution for medical image segmentation tasks.