Leukemia is a blood cancer that disrupts normal blood cell production by the uncontrolled growth of abnormal white blood cells. The classification of leukemia includes three classes: Acute Myeloid Leukemia (AML), Acute Lymphoblastic Leukemia (ALL), and nonleukemia (i.e., healthy sample). An accurate classification of leukemia is crucial for effective diagnosis and treatment of the disease. This research aims at proposing a deep learning based method for leukemia prediction and classification. We propose a reconfigured Weighted Convolutional Neural Network (W-CNN) based model to classify leukemia. In addition, we generate a new dataset for efficiently training the proposed model. The dataset were formed by exploiting proper gene expression microarray datasets and preprocessing them. Then, CatBoost with SHAP (i.e., SHapley Additive ex-Planations) were exploited to identify 33 important genes that most influence the leukemia classification to increase the classification effectiveness without relying on synthetic data augmentation. The generated dataset with identified genes, which consists of over 3,600 samples with 22,277 gene features, were used to train our proposed model. We conducted a number of experiments in various scenarios to evaluate our proposal. The experimental results show that our proposed method achieved 99.56% of accuracy, which is a promising result to be deployed for real-world classification of leukemia.