Automated detection and classification of leukemia on a subject-independent test dataset using deep transfer learning supported by Grad-CAM visualization. (May 2023)

• Record Type: Journal Article
• Authors: Abhishek, Arjun; Jha, Rajib Kumar; Sinha, Ruchi; Jha, Kamlesh
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Chronic lymphocytic leukemia -- Acute lymphoblastic leukemia -- Chronic myeloid leukemia -- Acute myeloid leukemia -- Deep transfer learning -- Gradient-weighted class activation mapping technique; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104722
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Leukemia is a type of cancer that affects blood cells and causes fatal infection and premature death. Modern technology enabled by the machine and advanced deep learning algorithms can assist doctors in diagnosing the disease in the early stage. Automatic classification of disease based on advanced deep learning always require a huge number of labeled dataset. This paper presents an image dataset of 750 microscopic blood smears. The dataset contains images of Chronic Lymphocytic Leukemia, Acute Lymphoblastic Leukemia, Chronic Myeloid Leukemia, and Acute Myeloid Leukemia. The dataset is enlarged by merging images of a dataset present in the literature, containing 500 images of Acute Lymphoblastic Leukemia, Acute Myeloid Leukemia, and normal cases. This dataset is used for automated detection and classification of leukemia using deep transfer learning, which is the major task of the proposed work. The impact of subject-independent test dataset on the robustness of trained models is also analyzed on different data splits for training and testing. For the classification task, the proposed work obtained an accuracy of 84% on a subject-independent test dataset when support vector machine is trained on features extracted by a pruned VGG16 whose last three convolutional layers are fine tuned. The features responsible for classifying an image to a particular class are visualized with the help of Gradient-weighted Class Activation Mapping technique. The dataset is being created by taking the opinion of various experts that will assist the scientific community in conducting medical research supported by machine learning models. Highlights: A new microscopic blood image dataset of Chronic Lymphocytic Leukemia, Acute Lymphoblastic Leukemia, Chronic Myeloid Leukemia, and Acute Myeloid Leukemia is proposed. The proposed dataset contains 750 images and is enlarged by combining a dataset containing 500 images present in the literature. The combined dataset now contains 1250 images and this dataset is used for the detection and classification of leukemia with the help of deep transfer learning method. The impact of the subject-independent test dataset on the robustness of trained models is analyzed. A subject-independent test dataset contains images that do not belong to subjects whose images are present in the training dataset. The classification results obtained in the study are analyzed qualitatively as well as quantitatively, and the features responsible for classifying an image to a particular class are visualized with the help of class-specific heatmaps.