Stochastic gradient descent optimisation for convolutional neural network for medical image segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Medical image segmentation faces challenges from hair artifacts and illumination variations.
• Accurate segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop a novel deep convolutional neural network (CNN) for segmenting medical images.
• To compare CNN architectures (U-net, FCN) and loss functions for improved segmentation accuracy.

Main Methods:

• A novel CNN-integrated methodology was developed for segmenting chest X-ray and dermoscopic images.
• The study compared U-net and FCN architectures using Jaccard distance and Binary-cross entropy loss functions.
• Optimized stochastic gradient descent with Nesterov momentum was employed.

Main Results:

• The proposed CNN model demonstrated superior performance compared to U-net and FCN architectures.
• Key performance metrics included sensitivity (0.9913), accuracy (0.9883), and Dice coefficient (0.0246).
• Post-processing with a threshold technique refined segmentation boundaries.

Conclusions:

• The developed CNN methodology offers enhanced efficiency for medical image segmentation.
• This approach improves diagnostic capabilities and treatment determination for various clinical conditions.
• The model effectively handles challenges like hair artifacts and illumination variations in medical imaging.