Template Credit: Adapted from a template made available by Dr. Jason Brownlee of Machine Learning Mastery.

SUMMARY: This project aims to construct a predictive model using a TensorFlow convolutional neural network (CNN) and document the end-to-end steps using a template. The Colorectal Cancer Histology dataset is a multi-class classification situation where we attempt to predict one of several (more than two) possible outcomes.

In this Take1 iteration, we will construct several simple three-layer CNN’s to model the dataset. We will use the best network as the baseline model for future modeling iterations.

INTRODUCTION: This data set represents a collection of textures in histological images of human colorectal cancer. All images are RGB, 0.495 µm per pixel, digitized with an Aperio ScanScope (Aperio/Leica biosystems), magnification 20x. The histological samples contain fully anonymized images of formalin-fixed paraffin-embedded human colorectal adenocarcinomas (primary tumors) from our pathology archive (Institute of Pathology, University Medical Center Mannheim, Heidelberg University, Mannheim, Germany).

ANALYSIS: The baseline model’s performance achieved an accuracy score of 98.25% after 15 epochs using the training dataset. After tuning the hyperparameters, the best model processed the validation dataset with an accuracy score of 84.00%.

CONCLUSION: In this iteration, the TensorFlow CNN model appeared to be suitable for modeling this dataset. We should consider experimenting with TensorFlow for further modeling.

Dataset Used: Colorectal Cancer Histology Dataset, Kather JN, Weis CA, Bianconi F, Melchers SM, Schad LR, Gaiser T, Marx A, Zollner F: Multi-class texture analysis in colorectal cancer histology (2016), Scientific Reports (in press)

Dataset ML Model: Multi-class image classification with numerical attributes

Dataset Reference: https://zenodo.org/record/53169#.XGZemKwzbmG

One potential source of performance benchmarks: https://www.kaggle.com/kmader/colorectal-histology-mnist

The HTML formatted report can be found here on GitHub.