GM FASST: General Method for Labeling Augmented Sub-sampled Images from a Small Data Set for Transfer Learning

Abstract

In this work, the use of transfer learning was explored to achieve the overall goal of classifying a small number (approx. 215 images) of Nonlinear Multiphoton Multimodal Microscopy Images of unstained oral cancer biopsies into three categories—healthy, inflammatory, and cancerous. This is achieved by first training a neural network model to detect basic histological components of human tissues, such as the stroma and mucosa, from a much larger (5000 images) Kaggle data set containing images of stained human colorectal cancer biopsies. Experiments were conducted to optimize the model’s architecture and hyperparameters, i.e., hidden layers, optimizers, and API callback functions used prior to retraining the model on a new and much smaller data set consisting of 215 Nonlinear Multiphoton Multimodal Microscopy Images. In addition to having different class labels, these images were acquired using an entirely different imaging and detection set up, and thus have different features than the Kaggle data set. In order to expand the limited size of the Nonlinear Multiphoton Multimodal Microscopy Images data set; tiling methods were used to sub-sample and augment the images though standard transforms such as rotation, scaling and mirroring. This research shows transfer learning and data set re-sampling can improve classification accuracy by 10% over training on the smaller data set alone.