Distributed computing methodology for training neural networks in an image-guided diagnostic application.

Area of Science:

• Medical Imaging
• Computer Science
• Artificial Intelligence

Background:

• Colonoscopy is crucial for detecting gastrointestinal lesions.
• Training neural networks for lesion detection requires substantial computational resources.
• Existing methods can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop an efficient distributed computing methodology for training neural networks.
• To enhance the speed and scalability of lesion detection model training.
• To enable the use of diverse computing architectures for neural network training.

Main Methods:

• Proposed a distributed computing methodology using a parallel virtual machine.
• Partitioned training datasets across multiple processors.
• Implemented distributed evaluation of error function and gradient values for neural network training.

Main Results:

• The methodology demonstrated considerable speedup in training neural networks.
• Performance gains were particularly notable with large network architectures and training sets.
• The approach proved effective for training neural networks using various learning methods.

Conclusions:

• The proposed distributed computing approach offers a scalable and efficient solution for training neural networks in medical imaging.
• Parallel virtual machine implementation significantly accelerates the training process.
• This methodology can facilitate the development of more sophisticated AI models for colonoscopy lesion detection.