Classify Images of Colon Cancer

Project Description

This project aims to develop a machine learning system that can classify histopathology images of colon cells. The system uses a modified version of the "CRCHistoPhenotypes" dataset for this task. The dataset consists of 27x27 RGB images of colon cells from 99 different patients and is used to perform two tasks:

• Classify images according to whether a given cell image represents a cancerous cell or not (isCancerous).
• Classify images according to cell type, such as: fibroblast, inflammatory, epithelial or others.

For the first 60 patients, the medical experts have provided labels for isCancerous and cell-type. However, for the remaining 39 patients, the medical experts have only provided labels for isCancerous.

Experiment Design

	Total instances	Fibroblast	Inflammatory	Epithelial	Miscellaneous	Normal	Cancerous
Main dataset
Train	6332	1208	1627	2610	887	3722	2610
Validation	1584	302	407	653	222	931	653
Test	1980	378	509	816	277	1164	816
Extra dataset
Train	8307	N/A	N/A	N/A	N/A	5915	2392
Validation	2077	N/A	N/A	N/A	N/A	1479	598

• For the first 60 patients (Main dataset), we split the data into train, validation and test sets with the 60/20/20 ratio. The models are trained and fine-tuned on the train and validation sets. The test set is reserved to evaluate how well the models perform on unseen data.
• For the remaining 39 patients (Extra dataset), we split them into train and validation sets with the 80/20 ratio. This dataset is later used to improve the performance on the cancer detection task by utilizing the extra isCancerous labels.

Model Evaluation

1. Baselines

	Training time (s)	Accuracy	Precision	Recall	F1
Cancer detection
VGG	70	0.8955	0.8955	0.8955	0.8951
Customized model	38	0.8914	0.8978	0.8914	0.8926
SC-CNN	49	0.8924	0.8971	0.8924	0.8934
RCCNet	258	0.9126	0.9137	0.9126	0.9129
Cell type classification
VGG	60	0.7288	0.8030	0.7288	0.7502
Customized model	24	0.7455	0.7464	0.7455	0.7429
Softmax CNN	64	0.7237	0.8189	0.7237	0.7621
RCCNet	159	0.7687	0.7635	0.7687	0.7626

For the baselines, we have researched 3 different architectures and propose a customized model of our own based on the VGG architecture. All models are trained in 50 epochs. If there is no improvment within 10 epochs, the training will stop to avoid overfitting.

The models use Adam optimizer with intial learning rate of 6 x 10^-5 iteratively decreased by a factor of \sqrt{0.1} if there is no improvement of validation loss during the training. The images go through an augmentation process where they are distorted, zoomed and flipped randomly.

2. Model improvement

Model	Training time (s)	Accuracy	Precision	Recall	F1
Cancer detection
RCCNet (main)	258	0.9126	0.9137	0.9126	0.9219
RCCNet (main + extra)	521	0.9040	0.9111	9.9040	0.9052
Cell type classification
RCCNet	159	0.7687	0.7635	0.7687	0.7626
RCCNet + pseudo labelling	399	0.8040	0.8033	0.8040	0.8032

We attempt to improve the performance on the cancer detection task by combining the training and validation sets from the main and extra datasets to retrain our model. The result on the test set is worse, which suggests there might be lots of noise on the extra dataset.

For the cell type classification task, we apply a technique called pseudo labelling with the following steps:

• Train a classification model on the main dataset (called the teacher model).
• Use the teacher model to predict cell types for the extra dataset.
• Combine the main dataset and the predicted labels on the extra dataset to train a new classification model (called the student model).

The result shows significant improvement on the test set, which indicates our method is effective.

3. Parameter tuning

Run	Batch size	Optimizer	Training time (s)	Accuracy	Precision	Recall	F1
Cancer detection
1	64	Adam	258	0.9126	0.9137	0.9126	0.9219
2	64	SGD	171	0.8909	0.8963	0.8909	0.8919
3	64	Nadam	181	0.9121	0.9130	0.9121	0.9124
4	32	Adam	169	0.9152	0.9162	0.9152	0.9154
5	32	SGD	152	0.8995	0.9015	0.8995	0.9000
6	32	Nadam	114	0.9116	0.9124	0.9116	0.9118
Cell type classification
1	64	Adam	399	0.8040	0.8033	0.8040	0.8032
2	64	SGD	331	0.7646	0.7854	0.7646	0.7727
3	64	Nadam	358	0.7944	0.7987	0.7944	0.7956
4	32	Adam	372	0.7960	0.7970	0.7960	0.7954
5	32	SGD	385	0.7798	0.7831	0.7798	0.7799
6	32	Nadam	420	0.8081	0.8136	0.8081	0.8099

At final stage, we try to tweak the batch size and optimizer to improve our models. The results show a slight improvement on both tasks.

Conclusion

At the end of the project, we delivered the two best models for the two tasks mentioned above after thorough research and analysis.

• Cancer detection: The best model is the RCCNet trained on the main dataset, using the Adam optimizer with batch size 32.
• Cell type classification: The best model is the RCCNet using the pseudo labelling technique. The model uses the Nadam optimizer with batch size 32.

For future work, we could perform a more in-depth process for parameter tuning by tweaking more parameters inside each layer such as padding, stride, number of filters, etc. to further improve the models' performances.

Acknowledgement

Thank you Jin for helping me deliver this project. You have been a great teammate!