Deep Learning Based Cancer Detection in Bone Marrow using Histopathological Images

• Published in: 2023 IEEE International Conference on Integrated Circuits and Communication Systems (ICICACS) pp. 1 - 8
• Main Authors: Rani B, Swaroopa, B, Geetha, G Shivaprasad Yadav, S, Shivakanth, Gandla, B M, Manjula
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.02.2023
• Subjects: Ash; Bones; Cells (biology); Clustering algorithms; cnn algorithm; gray level co-occurrence matrix; linear distinction stretching; Microscopy; plane extraction; PSNR; sickle cell anemia; White blood cells; world thresholding
• DOI: 10.1109/ICICACS57338.2023.10100116

The cancerous proliferation of abnormal white blood cells (WBC) in the bone marrow and blood is characteristic of leukemia, which has its origins in these blood-forming tissues. Blasts are immature aberrant cells that inhibit the creation of new, lineage-normal blood cells. Acute Lymphoblastic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myelomonocytic Leukemia (CML) are the four Chronic Lymphocytic Leukemia (CLL), Chronic myeloid leukemia (CML), and Acute myeloid leukemia (AML) (CML). Patients with leukemia are often diagnosed after their symptoms have been evaluated, a complete blood count (CBC) has been run, and a peripheral blood smear has been examined under a microscope by a pathologist. As part of the process to confirm and classify leukemia, a bone marrow examination and other sophisticated laboratory testing are performed. Traditional blood and bone marrow smear examination using light microscopy is plagued by intra- and inter-observer variability. Overcoming these individual biases of doctors is possible with the use of image processing-based approaches that can automatically analyse pictures of blood and bone marrow smears to detect abnormal cells. Using an American Society of Hematology dataset, the researchers classified cases of ALL into B-cell and T-cell subtypes, as defined by the World Health Organization (ASH). The research improves classification accuracy to 94.12% and compares the effectiveness of three distinct training strategies. The research also includes subtyping AML into M3, M4, and M5 using a ResNet-50 pretrained network and traditional machine learning based classifiers based on the CNN's characteristics. Using the features of ResNet-50, we were able to get a classification accuracy of 96.43 percent, and we were able to get a hundred percent accuracy using just a small number of traditional classifiers. Using the properties of Element and an ASH dataset, the researchers also sought to classify cases of chronic leukemia into CLL and CML. Classification using the Element yields an accuracy of 92.59%, while classification using the SVM with the Lucent's deep features yields an accuracy of 98.15%. The study's classifications are based on a straightforward framework that doesn't need sophisticated picture segmentation or feature extraction. The research has various limitations, one of which is that the dataset utilized in some of the classifications is not particularly big, which leads to a loss in classification accuracy. To address this, the research used data augmentation methods.