Evaluation of transfer learning on three medical image classification tasks

Abstract: Medical image diagnosing is a domain of medicine which has proven suitable for automation through machine learning. Automation would reduce cost and increase efficiency, making healthcare more widely available and decreasing the time needed to arrive at a diagnosis. One limiting aspect of this automation is the amount of available data needed for training. Transfer learning is a technique used to mitigate the negative effects of limited data. In this thesis the effects of using transfer learning on performance was examined on three different datasets: APTOS2019, ISIC2019, and PatchCamelyon. Results showed that on two of the three datasets (APTOS2019, ISIC2019), pretraining had a significant positive impact on performance. On the PatchCamelyon dataset, pretraining had a small positive impact on performance. The relatively small performance gain on PatchCamelyon can be understood by the large size of the dataset.