Histopathological-Cancer-Detection

Created an algorithm to identify metastatic cancer in small image patches taken from larger digital pathology scans. The data for this competition is a slightly modified version of the PatchCamelyon (PCam) benchmark dataset

Motivation

• Clinical diagnosis of breast cancer is best achieved by a biopsy. The pathologist diagnose by manually examining the histology slides under the microscope. However, the traditional diagnosis system needs expertise and only experienced pathologists can accurately determine the tumor tissues.
• Currently, in various rural parts of India, people do not have access to good healthcare facilities. Also, new advanced equipment are not available in rural areas so it is even possible that patients are not diagnosed properly. One of the primary reasons for dismal healthcare in rural areas is shortage of experienced doctors.

Dataset

The dataset used for the research is a slightly modified version of the PatchCamelyon (PCam). The original PCam dataset contains duplicate images due to its probabilistic sampling, however, this version does not contain duplicates. The dataset is open-source and can be downloaded from (https://www.kaggle.com/c/histopathologic-cancer-detection/data). The dataset has more than 220K RGB images with a dimension of 96x96x3. The given problem is the binary classification problem where the associated label has two class labels i.e. tumor and non-tumor tissues. A positive label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue.

Research Objectives

• Train various state-of-the-art deep learning architectures.
• Do a comparative study among the various models.
• Compare train-test splitting with stratified K-fold Cross-Validation.
• Apply image processing techniques such as Contrast Limited Adaptive Histogram Equalization (CLAHE).
• Apply various Image Augmentation techniques.
• Replace adaptive learning rate with one cycle policy(cyclic learning rate).
• Improve the model predictions with Test Time Augmentation(TTA).
• Measure the model performance on Accuracy, Precision, Recall, F1-measure and AUC ROC.

Prerequisites

• Python (3.6)
• Keras (2.3.0)
• Tensorflow (2.1)
• NumPy (1.14.3)
• Sklearn (0.22.2)
• Albumentations (0.4.5)
• pandas (0.22.0)
• tqdm (4.19.8)
• opencv-python (3.4.2)

How to Run Code

1. Git clone the repository to your local computer.
2. Download the data from Kaggle.
3. Open a terminal in the folder containing the repository.
4. Run python train_model.py for training a model using hold-out train-validation split.
5. Run python train_k_fold.py training a model using stratified K-Fold cross validation split.
6. Run python gradcam_vis.py for visualizing the gradient class-activation maps.

train_k_fold_tpu.py can only run either on Google Colab or GCP because TPUs are currently available on these online platforms.

Methodology Flowchart

The figure below shows flowchart of our proposed methedology:

Results Table

K-Fold result table of Model/baseline_model.py:

CNN Model	AUC-ROC
baseline_model with 4 Conv2D layers(w/o K-fold)	0.9348
baseline_model with 4 Conv2D layers(with 5-fold)	0.9458
baseline_model with 7 Conv2D layers(with 5-fold)	0.9643
baseline_model with 7 Conv2D layers(with 5x2-fold)	0.9666

Comparative result of SOTA models based on Train-Validation split of 90-10%:

Architectures	Accuracy	Precision	Recall	F1-score	AUC-ROC
DenseNet-169	0.98	0.98	0.98	0.98	0.9698
DenseNet-201	0.97	0.97	0.97	0.97	0.9700
SE-ResNet-50	0.968	0.97	0.97	0.97	0.9704
Inception-V3	0.9693	0.97	0.97	0.97	0.9655
SE-Inception-V3	0.9558	0.96	0.96	0.96	0.9681
Inception-V3 + CBAM	0.9656	0.97	0.97	0.97	0.9679
Xception	0.9821	0.98	0.98	0.98	0.9608
OctaveResNet-50	0.959	0.96	0.96	0.96	0.9709

Visualization

Following figure shows gradient class activation maps (Grad-CAM) on the penultimate convolutional block, which depict what areas of the image the CNN found most relevant for outcome prediction. The heat map represents how important each region of the image is to the given classification. This information can potentially be used by pathologists to make further hypotheses regarding the nature of the tumor.

License

This repository is licensed under the terms of the MIT license.