Hybrid handcrafted and deep feature fusion for automated acute myeloid leukemia classification using TCMA-Net on a class-balanced dataset.

Acute myeloid leukemia (AML) is a life-threatening hematological malignancy that requires accurate and timely diagnosis for effective clinical management. However, conventional cytomorphological analysis is time-consuming, subjective, and highly dependent on expert interpretation. Existing artificial intelligence-based approaches are often limited by severe class imbalance and inadequate feature representation, which restrict their generalization capability. To address these challenges, this study proposes a robust and high-performance AML classification framework based on hybrid feature fusion and deep learning. The proposed approach integrates handcrafted features with deep representations extracted from DenseNet201 and MobileNetV2 to form the Hybrid AML Descriptor Network (HADNet), which is further combined with the Transformer-based Tri block Convolutional Multi head Attention Network (TCMA-Net) classifier to capture both local and global contextual information. In addition, a Synthetic Acquisition Artifact Augmentation (S3A) strategy is introduced to mitigate class imbalance by generating realistic variations in illumination and noise. Extensive experiments were conducted on the AML cytomorphology Ludwig Maximilian University (LMU) dataset, where the proposed framework achieved a testing accuracy of 99.20%, outperforming traditional machine learning models and deep learning baselines. The effectiveness of the approach is further validated through ablation studies and external evaluation on the acute lymphoblastic leukemia (ALL) dataset, demonstrating strong robustness and generalization across different data distributions. Overall, the results indicate that the integration of handcrafted and deep features significantly enhances classification performance, stability, and reliability. The proposed framework provides a promising solution for automated AML diagnosis and has the potential to support clinical decision-making systems.