Deep DenseNet Convolution Generative Neural Network to Identify Leukemia

Abstract In today's global environment, there is a significant risk of cancer due to changing eating habits. Leukemia is a cancer affecting the bone marrow and white blood cells. This type of cancer affects children and adults, so it has a high mortality rate worldwide. This type of cancer invades the blood and spreads to other organs. The existing leukemia diagnosis system is manual, cumbersome, and prone to errors. Most existing systems fail to effectively analyze feature margins and relative entity thresholds, resulting in reduced precision. This occurs because actual margins of true positive rates degrade, increasing the false-positive rates. Hence, there is a need for an automated leukemia diagnosis system based on computer-aided diagnosis and deep learning methods to detect cancer cells from microscopic images. However, automatic leukemia diagnosis from microscopic blood images is tedious and produces imprecise, unreliable results. To address these problems, this research proposes a novel Deep DenseNet Convolution Generative Neural Network (D2CGNN) based on Absolute Maximum Support Feature Selection (AMSFS) to predict early-stage leukemia. The proposed first step is the Adaptive Gaussian Filter technique to analyze the noise ratio and remove noise from collected microscopic blood images. The second step is the Recursively Brightness Enhancement Histogram Equalization (RBEHE) method to improve image contrast levels from preprocessed images. The third step is identifying the region of interest in leukemia cancer using Region-based Watershed Segmentation. Afterward, select the best features of leukemia cancer using the AMSFS method. Finally, based on feature-selected microscopic images, the proposed method efficiently classifies samples as either cancerous or not. The proposed method was tested using the American Society of Hematology microscopic image, a publicly available database of leukemia patients. The proposed method provides better sensitivity of 93.4%, specificity of 93.2%, f1-score of 94.7%, and classification results of up to 95.6% than other methods.