Photonic Data Science

Nowadays, deep learning has been widely implemented into biomedical applications, but it is problematic to acquire large annotated medical datasets to train the models. As a technique for reusing knowledge obtained from one domain in another domain, transfer learning can be used with only small datasets. Despite of some current research about model transfer methods for medical images, it is still unclear how sample size influences the model performance. Therefore, this study focuses on the estimation of required sample size for a satisfactory performance, and also compares transfer methods with only 200 images randomly chosen from a colorectal cancer dataset. Firstly, based on a K-fold cross-validation, the balanced accuracies of 3 transfer learning networks (DenseNet121, InceptionV3 and MobileNetV2) were generated, and each network used 3 model transfer methods, respectively. Afterwards, by curve fitting with inverse power law, their learning curves were plotted. Furthermore, the estimation of required sample size as well as the prediction of final performance were calculated for each model. In addition, to investigate how many images are needed for curve fitting, the maximum number of images also changed from 200 to smaller numbers. As a result, it is shown that there is a trade-off between predicted final performance and estimated sample size, and suggested model transfer methods for large datasets do not automatically apply to small datasets. For small datasets, complicated networks are not recommended despite of high final performance, and simple transfer learning methods are more feasible for biomedical applications.