Classification of Microscopic Images

Acute Lymphoblastic Leukemia (ALL) is a type of blood cancer. The classification of B-lineage ALL based on microscopic images is a very challenging task due to high similarity of cells. This paper investigates problem of classification B-ALL and compares different approaches. Both classic and deep learning methods are used for classification of B-lineage acute lymphoblastic leukemia versus normal white blood cell from microscopic images. The subject level variability is shown for cancerous cells by adversarial validation procedure, whereas normal cells have similar appearances. Deep learning models based on ResNet architecture have significantly higher validation scores compared to gradient boosting model on hand crafted features, but also overfit heavily. Data set is provided by SBILab. The train data consist of 12528 images for 101 unique patients, 60 of them are cancerous (8491 images). One patient belongs only to one class. Each image has a resolution of (450, 450) pixels and one cell is in the middle of it, segmented from the background. Classic machine learning algorithms show their ability to solve the problem of identifying cancerous white blood cells with simple statistical features like color, texture and shape extracted from the images of cells. The scores are not so high, but using interpretative features can help to understand the data and explain how does a model makes decisions. This fact among with time and resource effectiveness provides standard models with an additional advantage over neural networks at the cost of quality loss.