Peripheral Blood Smear Images Classification for Acute Lymphoblastic Leukemia Diagnosis with an Improved Convolutional Neural Network

Abstract

Acute Lymphoblastic Leukemia (ALL) is one of the most common types of cancer globally, and the invasive tests used to diagnose it are costly and time-consuming. Peripheral Blood Smear (PBS) images are often preferred for the initial screening of cancer for the diagnosis of ALL. The manual search of PBS images for cancer diagnosis in laboratories is prone to errors due to certain factors such as interoperability errors and human fatigue. Moreover, there are serious problems in using trained data due to the non-specific nature of the signs and symptoms of blood smear images and the complex morphological structure. Advanced techniques have outstripped handcrafted and traditional methods for solving image processing problems. In this paper, an improved ResNet50 Convolutional Neural Network (CNN) model that uses a hybridization of Particle Swarm Optimization (PSO) to detect ALL and its subtypes, is proposed. The last 4 layers have been removed from the ResNet50 backbone model and 11 new layers have been added instead. The model is trained on an augmented dataset by applying color threshold-based segmentation in the Hue/saturation/value (HSV) color space to a publicly available dataset with 3256 PBS images. It has been observed that the performance obtained by classifying the trained improved ResNet50 model outperforms in terms of 99.65% accuracy. It can be said that the proposed method can help to distinguish between different categories of ALL and to determine the diagnosis and treatment protocol for laboratory workers.