Enhanced Leukemia Detection Based on Sequential Feature Selection with Optimized Random Nearest Neighbor Model for Microscopic Images

Abstract

The prevalence of leukemia as a life-threatening disease emphasizes the importance of early detection for effective treatment. The medical image processing performs an important role in the diagnosis of leukemia. Blood cancer that results from abnormal or not fully formed White Blood Cells (WBCs). ALL in the WBC are the body's warriors against infected cells. The abnormal proliferation of WBC in the bone marrow can damage other cells and affect bone marrow and lymphoid tissue. The manual morphological analysis of blood cells by hematologists is time-consuming, results depend on ineffective for large datasets, expertise, and can be biased. Blood cancers are also varied and complex in shape, texture, color and severity. In addition, different staining and light changes make the identification of blood cancers more difficult. To resolve the issue, a Random Nearest Neighbor (RNN) methodology based on Machine Learning (ML) protocol was deployed. To classify the blood cancer images, we deployed such techniques. In first, we collected a blood cancer dataset at Kaggle Website. To perform image processing the RNN contains four stages they are preprocessing, segmentation, feature selection, and classification. The initial phase is preprocessing, each pixel is replaced by a weighted average of its neighbors by the use of Bilateral Filter method. This improves the image quality and makes extracting meaningful features easier. Bilateral filters protect edges while smoothing the rest of the image. This helps to preserve the important features of the leukemia images. The second stage is segmentation, in this stage the images identifies and groups pixels or regions together based on similarity criteria based on Region Growing segmentation method. It starting from a seed point, adjacent pixels are added back into the region based on similar criteria. This criterion may be based on intensity values or other features relevant to the leukemia data set. After the segmentation we select an image feature through Sequential Feature Selection (SFS) method. This method can use it to remove unnecessary features from large datasets, or to progressively add and select useful features. Finally, we classify the images through RNN method to analyze dataset images of blood examples to categorize irregular cells symptomatic of blood cancer. The RNN method uses the concept of nearest neighbors to classify leukemia images, but also uses a random selection mechanism to improve classification accuracy of 94.5%. This method addresses this limitation by introducing randomness into the neighborhood selection. This helps to ensure reliable classification even in presence of noise. The investigational consequences demonstrate that our deployed methodology beats the existing model in accuracy, sensitivity, F1 score, specificity and error rate.