A deep transfer learning-based approaches for the detection and classification of acute lymphocytic leukemia using microscopic images

Authors

Ashish Kumar, Nishant Kumar, Jeril Kuriakose, Pushpendra Singh Sisodia

Citation (2024)

Kumar, A., Kumar, N., Kuriakose, J. et al. A deep transfer learning based approaches for the detection and classification of acute lymphocytic leukemia using microscopic images. Multimed Tools Appl (2024).

Abstract

Traditionally, researchers, doctors, and hematologists have faced difficulties in making a timely diagnosis of leukemia. Since, unlike other malignancies, leukemia typically does not create a (tumor) that can be seen in imaging techniques like X-rays or CT scans. As a result the procedures available at medical diagnosis centres of leukemia consumes a lot of time. Influenced by the capabilities of artificial intelligence techniques diagnosing the disease, the paper introduces advanced CNN models which includes InceptionV3, DenseNet201, Xception, ResNet152V2, and two hybrid models, i.e., InceptionResNetV2 and XceptionInceptionResNetV2 for identifying and classifying normal and leukemia cancer cells. Accuracy, root mean square error, recall, precision, F1 score, and loss are used to evaluate all applied advanced learning models. The Acute lymphoblastic leukemia dataset, which is separated into two classes: normal cells and malignant leukemia cells, is used in this study. During the pre-processing stage, every image undergo enhancement and are visually shown to fetch the color channels in the shape of an RGB histogram. Later the images are augmented to produce regions of interest by generating extreme points and employing adaptive thresholding techniques before being provided to the applied models for training. During the experimentation, it was discovered that InceptionResNetV2 had the highest validation accuracy of 98.59%. Similarly, DenseNet201 had the highest precision (97.57%), followed by InceptionV3 with recall (95.77%) and F1 score (95.56%). Moreover, the confusion matrix has also been generated to obtain the models’ recall, precision, and F1 score values for different classes of the dataset.

Keywords

Leukemia · Acute lymphoblastic leukaemia · Advanced CNN models · Deep learning · Hybrid model

View File Visit Publication