Contest Link: https://www.kaggle.com/c/syde522/overview Dataset Info: https://www.nature.com/articles/srep27988 Dataset Summary:

• This dataset contains segments of 10 Colorectal cancer tissue slides obtained from the pathology archive at the University Medical Center Mannheim
• These slides were then divided into 150x150 images in multiple different classes (from the link above):
  • 01_TUMOR - Tumour epithelium
  • 02_STROMA - Simple stroma (homogeneous composition, includes tumour stroma, extra-tumoural stroma and smooth muscle);
  • 03_COMPLEX - Complex stroma (containing single tumour cells and/or few immune cells)
  • 04_LYMPHO - Immune cells (including immune-cell conglomerates and sub-mucosal lymphoid follicles)
  • 05_DEBRIS - Debris (including necrosis, hemorrhage and mucus)
  • 06_MUCOSA - Normal mucosal glands
  • 07_ADIPOSE - Adipose tissue
  • 08_EMPTY - Background (no tissue)

Best performing Solution:

• Create 5 splits of the data
• Fine-tune a VGG16 network on each split
  • Training takes 3 steps. For each step, unfreeze a few more layers of the network
  • Part of the logic here was also that every time training is restarted, the optimizer is reset, possibly bumping the network out of some local minimmums
• Average predictions of each VGG16 network for final submission
• Results: 96.4% test set accuracy

File List:

• create_datasets.ipynb create 5 train/test combinations using sklearn's KFold
• vgg16_retraining.ipynb fine-tuning of a single VGG16 network using one fold of data
• create_ensemble.ipynb create ensemble network of multiple VGG16 models and use them to make predictions (this one is a bit messy)

Ideas for the Future (Partially from here)

• Incorporate Test time Augmentation to further improve model consistency
• Create a specific data augmentation function to simulate different concentrations of dye in the images from microscope slides
• Look into use of RAdam/Ranger for a more robust convergence: https://github.com/LiyuanLucasLiu/RAdam, https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
• Try fine-tuning Efficientnet pretrained with noisy-student method for better performance