Histopathological Cancer Detection

Among the many steps in cancer identification, one conclusive test involves sampling tissue from suspicious lumps and observing it under a microscope to ascertain if it is malign. Manual analysis of slides under a microscope is a painstaking and time-consuming task, often prone to human error. This project aims to explore machine learning and neural networks as a tool for accurate detection of metastatic cancer. We use the PatchCamelyon dataset which contains histopathologic scans of lymph node sections of size (96 x 96 px) which are labelled either 0/1 for absence/presence of cancer which we are trying to predict. We conducted experiments with three different deep learning models and performed data augmentation on the input dataset to add regularization, making it more robust to unseen data. We also performed test-time augmentation to further improve model efficiency. The CNN model from scratch and models built from ResNet50 and MobileNetV2 were trained with 3 different methods of setting learning rates across epochs:

• Constant learning rate: Fixed learning rate across all epochs.
• One Cycle Policy: Learning rate is initially gradually increased on a slope upto a maximum value, and then reduced. Extremely low learning rate applied in the final epochs.
• Reduce Learning Rate on Plateau: Learning rate reduced by a fixed factor when model’s validation accuracy does not improve over some epochs.

We have used TensorFlow and Keras API to implement all the models.

Logistic Regression and CNN:

1. Execute the following command in the execution directory: ln -s /datasets/home/91/391/sakundu/Project .
2. The provided notebook can be run step by step.

MobileNetV2:

1. Open the notebook on Google Colab.
2. The provided notebook can be run step by step.

ResNet50:

1. Update the variable data_base_dir to '/datasets/home/91/391/sakundu/Project'
2. Prepend the following lines before importing modules: import sys sys.path.append('/opt/conda/lib/python3.7/site-packages')
3. The provided notebook can be run step by step.

Peruse through our report for a detailed explaination on the dataset used, our approaches in tackling the challenge and analysis of each of our approaches.