Artifact detection for RADCURE

Convolutional neural net to detect artifact detection in the Radcure dataset images.

Preprocessing

• Padding images to 512x512x512 and downsampling to 256x256x256 by default.

Augmentation

• Horizontal mirroring of images

Sampling

• Weighted Random Sampler with weights equal to the inverse of the class size

Neural Net

• Standard 3d Convolution layers with Batch Normalizatin and Maxpooling except for the las tlayer which average pools
• Fully connected layers with sigmoid output
• Using Leaky-Relu as the activation function
• Binary Cross Entropy as loss function
• using stochastic gradient descent with a step scheduler

Results can be found Here

Requirements:

• Python
• Pytorch
• Numpy
• Sklearn
• Skimage