A Comparative Study of Transfer Learning Networks and Siamese Networks for Acute Lymphoblastic Leukemia (ALL) Diagnosis

Abstract

Acute Lymphoblastic Leukemia (ALL) is a type of blood cancer that primarily affects the white blood cells. ALL is a serious problem in our society, accounting for approximately 25% of all pediatric cancers. Another major issue of ALL is that the healthy and cancerous cells look extremely similar to the human eye, leading to high chance of misdiagnosis. A novel method was proposed to solve a two-fold problem using Few Shot Learning (FSL), a machine technique that uses a small support set to perform binary classification. Traditional machine learning methods require extremely large medical training datasets, which can lead to high computation expenses and are difficult to access due to patient privacy. Convolutional neural networks (CNNs) such as VGG19 and GoogLeNet were also compared to investigate if there was any significant improvement with FSL on the CNMC dataset. Currently, our Siamese Network has a 85% testing accuracy when training it on 10 Epochs. However, transfer learning has also shown to be a way to use a small amount of data to receive a high accuracy rate, and require a less computationally intensive training process.An accuracy rate of 99% and 96% was achieved on GoogLeNet and on VGG-19, which were trained for 50 epochs with image transformation preprocessing. Overall, traditional CNNs continue to outperform FSL methods, but methods are still a viable option for second opinions.