A HYBRID CNN-GNN FRAMEWORK FOR FINE-GRAINED PERIPHERAL BLOOD CELL MORPHOLOGY CLASSIFICATION

Abstract

Identifying blood cell types correctly under a microscope is a task doctors rely on constantly, since red cells, white cells, and platelets each carry information about a patient’s health, and several conditions leave visible traces in how these cells look. Doing this by hand takes time, varies from one observer to another, and needs specialists who are not always avail- able, especially in smaller clinics. In this paper we build a Hybrid CNN-GNN model that tries to combine the best of two worlds: a pretrained ResNet18 branch that picks up on color and texture patterns, and a graph neural network branch that represents each cell as three connected parts, the cell it- self, the nucleus, and the cytoplasm, so the model can also reason about how these parts relate to one another. We train and test the approach on the OI-PBC dataset, which has close to 300,000 labeled peripheral blood cell images spread across twelve cell types, and we balance the training data to 3,000 images per class so that common cell types do not drown out the rarer ones. On a test set that is also balanced this way, our hybrid model reaches 91.15% accuracy and a macro F1 of 91.17%, ahead of a GNN-only version (48.42%) and a CNN- only version (90.38%). On the original, unbalanced test set it hits 99.35% accuracy, a Kappa of 0.966, and an AUC of 0.990. The task here is twelve-class cell-type classification, not disease diagnosis, since our labels describe cell morphology rather than confirmed patient outcomes. We also wrapped the pipeline in a small Gradio research prototype that returns a predicted cell type, a confidence score, and a downloadable PDF summary, and we deliberately keep this framed as an exploratory screening aid rather than a validated diagnostic tool, since that would require patient-level labels and clinical validation we do not yet have.