Development and Evaluation of an Explainable Deep Learning Prototype for 13-Class Peripheral Blood Cell Classification

Note: The repository contains preliminary research and code. It has not undergone formal peer review. Manual differential counting of peripheral blood smears is a time-consuming and subjectively variable task in clinical hematology. This project details the development, evaluation, and deployment of an explainable deep learning prototype for the classification of 13 peripheral blood cell types. Utilizing the KU-Optofil Peripheral Blood Cell (PBC) dataset (31,489 images), a lightweight MobileNetV2 architecture was fine-tuned using mixed-precision training to accommodate limited computational hardware (4 GB VRAM), offering an edge-computing alternative to larger models. To address severe class imbalance and prevent gradient instability, a pipeline incorporating a WeightedRandomSampler was utilized rather than the class-weighted loss functions employed in prior approaches. Evaluated under the dataset's recommended strict patient-level separation, the model achieved a Test Set Accuracy of 94.31% and a Macro-F1 Score of 0.7041. This resource-constrained pipeline achieved comparable overall accuracy and substantially improved recognition of rare classes compared to the published DenseNet-121 benchmark, although the individual contribution of the sampling strategy was not isolated in a controlled ablation study. Error analysis via a normalized confusion matrix revealed that misclassifications aligned with known biological maturation continua. Gradient-weighted Class Activation Mapping (Grad-CAM) supported the interpretation that the model relied primarily on cellular morphology, though occasional contextual artifacts were identified. Robustness checks demonstrated high resilience to image perturbations, with a notable vulnerability to Gaussian blur that is consistent with a dependence on fine-grained morphological detail. The system is deployed as an educational web application featuring a FastAPI backend and a deterministic, rule-based explanation engine. This project demonstrates the viability of building transparent and computationally efficient AI prototypes for hematological education. Files included in the Github repository linked with the paper: Model weights and architecture configurations Training, evaluation, and explainability scripts (Grad-CAM) Web application deployment code