Deep Learning Based Framework for Detection and Classification of Leukemia Using Microscopic Images.

Abstract

Excessive proliferation of white blood cells (WBCs) in the bone marrow leads to a type of blood cancer known as leukemia. This blood cancer impairs the immune response, and timely detection and diagnosis are crucial for human health. Several manual and automated methods for leukemia diagnosis have emerged recently, with the latter still requiring medical practitioners' attention for leukemia treatment. Microscopy is an essential technique in the diagnosis of leukemia, as it allows examination of blood cells in detail and accurate identification of cancerous cells. But artificial intelligence (AI), especially deep learning, has recently been explored to enhance leukemia detection and classification. This study presents an approach to detect leukemia from microscopic images using a convolutional neural network (CNN). The approach starts with image pre-processing, then enhances the training dataset through data augmentation strategies, increasing the number of image samples from 270 to 1268. The U-Net model is used to segment leukemia and normal cells, allowing for efficient feature extraction from high resolution microscopic images. Then, the ASH dataset is classified using a CNN-based architecture, differentiating between the different subtypes of leukemia. Through 10-fold cross-validation, the proposed model achieves an accuracy rate of 99.06% for binary classification and 98.68% for multi-class classification, with a recall of 96.74%, a precision of 96.83%, and an F1-score of 96.77% ± 1.09%. These results suggest that the proposed model performs similarly to other methods. The framework of microscopic imaging and deep learning in our model shows promise as computer-aided diagnosis of leukemia. But more work is needed to train it on a larger and more diverse dataset. This process can be extended to detect other blood disorders through the inclusion of other deep learning models or potentially investigate robust data augmentation techniques in the future.