Advanced Deep Learning Framework for Accurate Detection and Classification of Bone Marrow Malignancies using YOLOv8

The diagnosis of bone marrow malignancies, including leukemia, lymphoma, and myeloma, has so far primarily depended on subtle variations in cellular morphology and tissue features, which pose a formidable challenge. To tackle these tipping points, the study proposes an advanced deep learning framework that exploits the features of the YOLOv8 model architecture for accurate detection and classification of malignancies in microscopic images of bone marrow. The modules proposed in the framework encompass three critical components: detection, classification, and evaluation. Detection, which is the very backbone of the YOLOv8 model, performs the task of real-time detection of regions of interest (ROIs) containing malignant cells using state-of-the-art CNN architecture. Then ROIs identified by these models are passed on to the classification module, where they will be used to classify the cells into malignancies versus non-malignancies. Using robust augmentation techniques like rotational transformations as well as edge detection allows the model to be robust via establishing orientation invariance and highlighting important structural features. Next, feature selection methods are applied, mainly SelectKBest and the Chi-square test to enhance computational efficiency by rejecting the least diagnostically important features. The proposed framework shows great promise, yielding an average AUC value of 0.9490±0.0124 across validation folds. A precision-recall curve and ROC analyses confirm the high sensitivity and specificity of the framework, hence minimizing false positives and false negatives. Thus, its capability for accurate and efficient detection of malignant tissues enables real-time use in clinics. With its capability of enhancing diagnostic accuracy and circumventing observer bias, the proposed system endows pathologists with a radical support tool for timely intervention.