Data-Driven Deep Learning Approach for Early Detection of Acute Lymphoblastic Leukemia

Abstract—Acute Lymphocytic Leukemia (ALL) is a bloodcancer that is life-threatening and in which timely diagnosis is important in enhancing survival chances of patients. Conven-tional ways of diagnosis process is based on manual examination of microscopes which are time consuming and subject to humanerror. The article suggests a new deep learning model namedQuality-enhanced Convolutional Residual Network (QCResNet)to perform automated multi-classification of ALL to four phases:Benign, Early, Pre, and Pro leukemia. The model also has residuallearning, batch normalization and dropout to maximize featureextraction and generalization. The issue of data imbalance issolved with a powerful preprocessing pipeline, consisting of pixelnormalization, stratified splitting, and class weights. C-NMCLeukemia dataset 19 experimental analysis using more than6,500 microscopic images of training, validation, and use has atraining accuracy of 98.27%, validity accuracy of 96.01%, andtest accuracy of 98.60%. The Pro-stage stage has 100 percentrecall with the guarantee of the good detection of advancedcases. The trained model is implemented in the form of aweb application based on Django, which offers real-time clinicaldecision support to detect early leukemia.