Next-Generation Architectures for Leukemia Subtype Identification: A Review of CNN, Vision Transformers, and Hybrid Models

One of the most urgent issues is the possibility to provide reliable and repeatable leukemia subtype diagnosis based on the hematological image due to the minimal morphological variations, inter-rater error, and due to the disadvantages of the traditional diagnostic procedures. These dilemmas are viewed in this review by conducting a systematic examination of next-generation deep learning architectures, including convolutional neural networks (CNNs), Vision Transformers (ViTs) and novel hybrid CNN-transformer models to automatically identify the subtypes of leukemia. The paper is an overview of the current advances in feature extraction, attention and multi-scale representation learning and how hybrid architecture can effectively combine local spatial feature learning and global contextual learning. Important results have shown that transformer-based and hybrid strategies are always better than the traditional CNN models in accuracy, robustness, and generalization on different datasets, and they also allow better interpretability when used in combination with explainable AI methods. In addition, the review also identifies the significant gaps in the research related to the heterogeneity of data, complexity of computation, and clinical validation. These are the lessons that suggest how the next-generation architecture will enhance the scalability and reliability of the diagnostics. Its results can be particularly relevant in the framework of developing AI-based clinical decision-support systems, which would enable an earlier diagnosis, less subjective diagnosis, and implementation in a healthcare environment with scarce resources.