A ROBUST DEEP LEARNING–BASED COMPUTER-AIDED DIAGNOSTIC FRAMEWORK FOR BLOOD CANCER DETECTION USING ADAPTIVE EFFICIENTNET OPTIMIZATION

Abstract

Leukemia is a malignant disorder of the blood-forming tissues characterized by uncontrolled proliferation of abnormal white blood cells, leading to impaired immune function and disruption of normal hematopoiesis. It represents a significant global health burden and accounts for a substantial proportion of pediatric cancers. Acute lymphoblastic leukemia (ALL), the most prevalent subtype, originates in the bone marrow and progresses rapidly, making early diagnosis essential for improving survival outcomes. To facilitate timely and accurate detection, this study presents an automated computer-aided diagnosis framework based on the EfficientNet-B3 convolutional neural network, integrated with a dynamically adjusted learning rate strategy. The proposed approach adaptively modifies the learning rate at each training epoch by jointly analyzing training accuracy and loss values to enhance convergence and classification performance. The model was evaluated on the research Leukemia dataset following normalization and class balancing, achieving average precision, recall, specificity, accuracy, and Dice similarity coefficient values of 98.29%, 97.83%, 97.82%, 98.31%, and 98.05%, respectively. The proposed approach employs a dynamic learning rate adjustment mechanism that continuously monitors validation loss and training accuracy to enhance convergence stability and classification reliability. Experiments were conducted on the public Acute Lymphoblastic Leukaemia dataset to evaluate cross-disease generalization. These findings confirm the effectiveness of the proposed method for automated ALL detection.