LeuFeatx: Deep learning-based feature extractor for the diagnosis of acute leukemia from microscopic images of peripheral blood smear

The abnormal growth of leukocytes causes hematologic malignancies such as leukemia. The clinical assessment methods for the diagnosis of the disease are labor-intensive and time-consuming. Image-based automated diagnostic systems can be of great help in the decision-making process for leukemia detection. A feature-dependent, intrinsic, reliable classifier is a critical component in building such a diagnostic system. However, the identification of vital and relevant features is a challenging task in the classification workflow. The proposed work presents a novel two-step methodology for the robust classification of leukocytes for leukemia diagnosis by building a VGG16-adapted fine-tuned feature-extractor model, termed as "LeuFeatx," which plays a critical role in the accurate classification of leukocytes. LeuFeatx was found to be capable of extracting notable leukocyte features using microscopic single-cell leukocyte images. The filters and learned features are visualized and compared with base VGG16 model features. Independent classification experiments using three public benchmark leukocyte datasets were conducted to assess the effectiveness of extracted features with the proposed LeuFeatx model. Multiclass classifiers trained using LeuFeatx deep features achieved higher precision and sensitivity for seven leukocyte subtypes compared to the latest research on the AML Morphological dataset, and it achieved higher sensitivity for all cell types vis-à-vis recent work on peripheral blood cells dataset from the Hospital Clinic of Barcelona. In a binary classification experiment using the ALL_IDB2 dataset, classifiers trained using LeuFeatx deep features achieved an accuracy of 96.15%, which is better than the other state-of-the-art methods reported in the literature. Thus, the higher performance of the classifiers across observed comparison metrics establishes the relevance of the extracted features and the overall robustness of the proposed model.