Implementation of pretrained CNNs on cancer detection
Haolin Li · 2022 IEEE 5th International Conference on Information Systems and Computer Aided Education (ICISCAE) · 2022
As one of the most universal and lethal cancers, Lung cancer results in millions of annual deaths in the United States alone. Though deadly, early and active interventions can remarkably improve a patient's survival rate in five years. Doctors need accurate information about the specific type of lung cancer, which can be time-consuming and sometimes hard to discriminate based on human resources alone. Deep learning applications in Computer Vision have recently achieved significant results, bringing possibilities for applying such techniques to cancer detection. Pretrained models have been proved to be efficient over image feature extractions. However, it is still unclear which of them suits lung cancer detection the most. Thus, comparison of performances between different pre-trained models is necessary. This paper explored performances across different pre-trained feature extractors followed by a fully-connected classifier. Furthermore, this paper also tested models over 30% of the original data to identify performances over the limited volume of data. Lastly, similarity calculations of output features based on both cosine similarities and T-SNE were performed. The results show that all models over full-sized data reached accuracy over 99%, some of which reached 99.8%. Performances over the limited volume of data are still remarkable, all of which are around 97%~98%. More importantly, though orthorhombic regarding cosine similarities, features from different extractors show apparent clustering behavior. This finding shows a consistent direction of learning across models, which brings new information about the interpretability of feature extractors in this field. Such positive outcomes further prevail the bright future of assisting cancer detection with deep learning.