Interpretable deep learning-based classification of polycythemia vera, secondary erythrocytosis, and healthy controls via tabular-to-color image transformation of routine blood parameters

Abstract Polycythemia Vera (PV) and Secondary Erythrocytosis (SE) are hematological conditions with overlapping clinical features, particularly elevated hemoglobin and hematocrit levels, which complicate differential diagnosis in routine practice. This study proposes a novel tabular-to-image deep learning framework for the classification of PV, SE, and Healthy Controls (HC) using routine hematological data. A total of 345 clinical samples (111 PV, 143 SE, and 91 HC) described by 28 hematological variables were processed through a multi-stage pipeline including MICE-based missing value imputation, feature engineering, mRMR feature selection, and deterministic HSV-to-RGB encoding of 36 selected features into 6 × 6 color-mapped grid images. Six state-of-the-art architectures (DINOv2, EfficientNetV2, ConvNeXt, Swin Transformer, ResNeXt-50, and Vision Transformer) were fine-tuned using transfer learning and Bayesian hyperparameter optimization with Optuna. Performance was evaluated using stratified 10-fold cross-validation, and model interpretability was examined with Grad-CAM and SHAP analyses. ResNeXt-50 achieved the best performance, with a macro F1-score of 0.9250 ± 0.0356 and an accuracy of 92.3%, while all models exceeded 91% accuracy. SHAP and Grad-CAM findings indicated that the models relied on clinically meaningful hematological features. These results demonstrate that color-mapped image representations of routine blood parameters provide an effective and interpretable basis for automated erythrocytosis classification, with potential utility as a computer-aided diagnostic tool in clinical hematology.