Determining critical lung cancer subtypes from gigapixel multi-scale whole slide H&E stains images

Being one of the world's deadliest cancers, lung cancer leads to the death of many in their first 5 years of diagnosis [22]. However, with early prognosis, the magnitude of the situation could be lessened and further mortalities could be prevented. The implementation of machine learning for medical assistance in clinical practice includes using computer vision for cancer diagnosis. In this paper, we utilized four different Convolutional Neural Network models: VGG16, Densenetl21, Resnet50 and Inceptionv3. All four models were pre-trained from the benchmark dataset ImageNet. We performed both fine-tuning and feature extraction for each model. For the fine-tuning, we utilized a self-defined train function that keeps track of the performance of the model after each 20 epochs. The best performing model would be saved locally for later use. For the feature extraction, we utilized a feature-based extraction and 3 different classifiers including Linear-SVM, Ridge Classifier and logistic regression. The results of the experiment demonstrated that the fine-tuning method had significantly better performance as opposed to the representation learning method. However, the fine-tuning method took a much longer time to train and test. As an extension, we incorporated level 0 resolution of the svs data files in an attempt to enhance the performance of the representation learning method by diversifying our data. Due to computational limitations however, we reduced the dataset to 1/4 of what it was before in order to incorporate level 0 resolution images. The performance with a more diversified data increased by 1-2%.