Acute Leukaemia Diagnosis Using Transfer Learning on Resnet-50

This paper presents an automated model for leukaemia detection that is based on the computational power of a deep pre-trained model Resnet-50. The conventional manual method to detect the disease from microscopic blood cell images is time driven and the diagnosis is subjective due to the variation of technical expertise of the hematologists and may vary from one pathologist to other. Hence a model is proposed that exploits the transfer learning technique on Resnet-50 to learn the features of microscopic blood cell images from the Acute Lymphoblastic Leukemia Image Database for Image Processing (ALL-IDB1) to classify them into diseased and healthy. As the number of images in the dataset is very less for training on deep-network, the model may overfit. As a precautionary measure, augmentation of images is performed during the training. Apart from image augmentation, L2 regularization is also used to reduce overfitting. The proposed model demonstrates 100% accuracy on unseen test images with Resnet50. The comparison of the obtained results is done with state-of-the-art work performed by contemporary researchers.