A deep ensemble medical image segmentation with novel sampling method and loss function.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Medical image segmentation is crucial for disease diagnosis and treatment planning.
• Deep learning models show promise but face challenges like class imbalance and accuracy.
• Existing methods struggle with precise identification of abnormal tissues and background.

Purpose of the Study:

• To propose a novel semantic segmentation approach for medical images.
• To address class imbalance and enhance segmentation accuracy.
• To improve the identification of regions of interest in medical scans.

Main Methods:

• A new sampling method to handle class imbalance in medical datasets.
• A novel pixel-level loss function inspired by exponential loss.
• An ensemble model combining two UNet models with ResNet backbones, trained on primary and sampled datasets.

Main Results:

• The proposed approach effectively handles class imbalance.
• The novel loss function and ensemble model enhance segmentation performance.
• Evaluated on Kvasir-SEG, FLAIR MRI LGG, and ISIC 2018 datasets, outperforming existing methods.

Conclusions:

• The novel sampling method and loss function improve medical image segmentation.
• The ensemble deep learning model offers a robust solution for accurate disease diagnosis.
• This approach advances the field of medical image analysis and computer-aided diagnosis.