Ensemble Deep Learning Models for Lung Cancer Diagnosis in Histopathological Images

Abstract

Computed Tomography, widely known as CT scan, is used for the detection of lung cancers. However, sometimes doctors cannot detect such harmful diseases based on the CT scan images; hence they recommend a manual biopsy. A manual biopsy is very tedious and exhausting and can also consume much time, and there is a chance of medical error in it. So, it is high time for researchers to develop a Computer-Aided Diagnosis (CAD) System that can help detect lung cancer in minimal time and effort with the highest accuracy. Lung biopsy reports are collected, and histopathological images are generated out of it. Image processing and data augmentation techniques are applied upon a collected histopathological image to prepare a dataset. Three different models are trained by using this dataset, Convolutional Neural Networks and its variants- Inception-V3 and ResNet50. These developed models suffer from high variance on unseen data as they are measured on training and validation datasets. As there is no obvious indication of which model is better than another toward the conclusion of the training period, picking the final model is dangerous based on the results. Ensemble learning techniques help to reduce high variance and give better accuracy than a single model. Ensemble learning provides many strategies to overcome such problems. This research demonstrates approaches to ensemble deep learning models to reduce high variance and improve predictive accuracy on unseen data.