DeepImageTranslator: A free, user-friendly graphical interface for image translation using deep-learning and its...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning has advanced image translation but often requires programming expertise.
• Existing graphical tools lack full customization for deep learning pipelines.

Purpose of the Study:

• To develop a user-friendly graphical interface for creating, training, and evaluating custom deep learning models for image translation.
• To assess the tool's effectiveness in CT image semantic segmentation and noise reduction.

Main Methods:

• Developed DeepImageTranslator with a Tkinter GUI and backend using Pillow, Numpy, OpenCV, Augmentor, Tensorflow, and Keras.
• Trained convolutional neural networks (CNNs) and evaluated data augmentation, deep-supervision, and sample size effects.
• Enabled customization of CNN architecture, optimizer, loss function, and augmentation schemes.

Main Results:

• Achieved state-of-the-art accuracy and generalizability in semantic segmentation and noise reduction.
• Demonstrated that highly accurate 3D segmentation models for body composition can be developed with as few as 17 training images.
• Showcased DeepImageTranslator's capability to customize all deep learning pipeline aspects.

Conclusions:

• An open-source, user-friendly deep learning tool for accurate image translation was developed and validated.
• DeepImageTranslator empowers researchers without programming experience to leverage advanced image translation techniques.
• The software is available for download, promoting accessibility and further research.