Colon cancer patch classification(benign vs Cancer) project using Dense-net.(2019.01~2019.06)

In this study, we propose a convolutional neural networks (CNN) for classifying colon dataset using Densely connected network. With changing the depth of the model and input size, we compared the model performance. As a result, the best model achieved 98.02% accuracy and 0.9951 AUC. Plus, as deeper depth, smaller input image decreased model performance.

• Size of Input to Gloval Avg.Pooling Layer

• Affine : Random scale, translation, rotation, shear.
• Horizontal and vertical flipping
• Texture : Gaussian blur, Median blur and Gaussian noise
• Color : Adding Hue, saturation, linear contrast

• Adam optimizer, default parameters
• 3 fold cross validation
• 40 epochs
• Cross-entropy loss

• First fold *
  • 
• Second fold *
  • 
• Third fold *
  • 

• H&E stained colon pathology images
• 1171 benign patches, 2472 tumor patches
• Among tumor patches, well differentiated 300, moderately differentiated 1701, poorly differentiated 471
• Provided by Kangbuk Samsung Hospital

• pytorch (version : 1.1, gpu-version)
• imaug - for image augmentation
• sklearn - for K-fold cross validation, dividing train/val datasets
• tensorboardX
• matplotlib
• numpy