Graph-Aware Medical Image Intelligence: A Comparative Study of CNNs, GNNs, and Multimodal LLMs for Acute Lymphoblastic Leukaemia Subtyping

Abstract The correct determination of the cell type of an acute leukaemia (ALL) is essential, since therapy depends on the morphological variations that can be seen in blood smear images. Manual microscopy is still a time-consuming, labor-intensive method and the results of the diagnosis often differ from one observer to another, which is the reason for the need of reliable automatic diagnostic models. This study compares and evaluates three deep learning paradigms - Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and multimodal Large Language Models (LLMs) - for the classification of four ALL subtypes from the Multi Cancer Dataset (4000 images). The CNN models used were trained with the same preprocessing steps, including resizing, normalization and augmentation, to promote generalization to other tasks. Based on the spatially segmented nuclei and morphological features, four GNN models were designed to capture inter-cell relationships that are not captured by pixel-based architectures. Furthermore, a multimodal model ViT–BERT and its fine-tuned version were also used to incorporate visual embeddings and text domain knowledge. The tuned ViT–BERT model achieved F1 = 0.99, matching the top CNN baselines (DenseNet121, MobileNetV2 at F1 = 0.98) and outperforming InceptionV3 and ResNet50. The graph-based reasoning models were found to perform competitively on the subtypes with unique spatial arrangement of cells, showing the importance of reasoning on graphs. Interpretability was addressed through Grad-CAM for CNN and multimodal models, and GNN Explainer for graph architectures, being transparent in model decisions. The results suggest that the pixel-level, graph-level, and multimodal representations are effective in enhancing accuracy and interpretability, which helps to build more reliable diagnostic systems for ALL.