From cell images to clinical insight: Clinically grounded explainable classification of peripheral blood and bone marrow cells with cross-modal learning

White blood cells (WBCs) are essential to immune defense, protecting against infection and disease. Peripheral blood smear examinations are commonly used to assess WBC morphology and diagnose haematological disorders. Manual identification of WBCs under a microscope is time-intensive and prone to human error. Advances in machine learning and deep learning have enabled automated WBC classification, with deep learning models demonstrating expert-level performance. Despite their effectiveness, the lack of clinically relevant interpretability remains a significant barrier to clinical adoption, as clinicians may be reluctant to trust the decision-making processes of black-box models. To address this gap, we present an Explainable Cross-Modal Framework for White Blood Cell classification (EXCF-WBC) that leverages contrastive learning as the core mechanism for achieving interpretability. EXCF-WBC aligns interpretable haematological features with deep visual embeddings in a shared latent space via dual projection heads, thereby grounding visual embeddings in clinically meaningful semantics. A large language model module translates the aligned multimodal evidence (image data, clinical features, and feature attribution values) into structured natural-language explanations, providing a transparent, evidence-based bridge from visual data to clinical interpretation. Quantitative faithfulness evaluation using insertion-deletion analysis demonstrates that clinical feature attributions achieve lower deletion AUC and higher insertion AUC than pixel-level Grad-CAM, indicating stronger alignment with model decision-making. A structured user study with 12 clinicians further confirms that the framework aligns with clinicians' reasoning (mean rating of 4.46/5). EXCF-WBC addresses the interpretability gap by delivering clinically grounded and trustworthy explanations while maintaining strong classification performance across proprietary and public datasets.