Hematologic Cancer Cell Detection and Classification Using Optimized VGG-19 with Stratified K-fold Cross-Validation

Abstract

The most prevalent pediatric blood malignancy is acute lymphoblastic leukemia (ALL). ALL is a lethal disease in which patients have a lower survival rate. Its prompt detection and precise categorization are essential for successful treatment.  Manual microscopic diagnosis is laborious, prone to mistakes, and heavily reliant on specialized knowledge.  With stratified 7-fold cross-validation, which assures an equal ratio of normal and malignant cells in each fold, this research provides an enhanced VGG-19-based deep learning model for reliable binary categorization of leukemic vs normal cells to overcome the drawbacks of the manual detection procedure. For this study, the C-NMC leukemia dataset used comprises single-cell images of normal (HEM) and cancerous (ALL) types. For the binary classification test, transfer learning was utilized by keeping the initial convolutional layers of the pre-trained VGG-19 model and swapping out its last few layers. This resulted in an accuracy of 98.87%, a sensitivity of 98.97%, 98.82% specificity, 99.68% precision, and an F1-Score of 99.23%.  The outcomes demonstrate how well the model handles morphological differences and class imbalance in leukemic cell images. In addition, the proposed model also outperformed the other pre-trained neural networks, viz., ResNet-18, ShuffleNet, and GoogleNet, in terms of accuracy.