A hybrid training method based on deep learning for medical images classification

Deep learning has gained remarkable results in many fields. However, using deep learning in medical fields is faced with the scarcity of datasets and the expensive cost of annotation data. So, this paper aims to demonstrate an exploration of whether different types of datasets can complement each other to fulfill the required number for training a model in deep learning. The first dataset is from the Kaggle, a modified version of the Patch Camelyon benchmark. Moreover, the second dataset is from the “Lung and Colon Cancer Histopathological Image Dataset,” which has 750 images of lung tissue and 500 images of colon tissues. The paper divides the datasets into five groups. It uses CNN and Resnet-18 separately to analyze and acquire each group's loss and accuracy value. According to the result of the experiment, using the Resnet-18 model can obtain a higher accuracy value than using the Custom-CNN model, and adding more datasets to the target dataset does not significantly influence the accuracy of the training model. Considering all datasets are imported directly, adding different related datasets by the specific method maybe is a proper way to solve the problem of limited amounts of data.