DeepImageTranslator

DeepImageTranslator performs image-to-image translation using convolutional neural networks (CNNs) to enable semantic segmentation and noise reduction of CT images for bioimage analysis.

Key Features:

• Customization: Customizable CNN architectures, training optimizers, loss functions, and image augmentation schemes.
• Computational Libraries: Back-end computations use Pillow, NumPy, OpenCV, Augmentor, TensorFlow, and Keras.
• Supported Tasks: Implements semantic segmentation and noise reduction (denoising) on CT images, including 3D CT segmentation for body composition analysis.
• Low-sample Performance: Demonstrates capability to produce accurate 3D segmentation models for body composition analysis with as few as 17 training images.

Scientific Applications:

• Semantic Segmentation: Segmentation of complex structures in 3D CT scans, applicable to body composition analysis.
• Noise Reduction: Reduction of noise in CT images to enhance image quality for downstream quantitative analysis.

Methodology:

Convolutional neural networks are trained and evaluated using a three-way cross-validation process. The effects of data augmentation techniques, deep supervision strategies, and varying sample sizes on model performance are systematically assessed.