MorphX: A WHO-Aligned Ante-Hoc Explainable AI Framework for Acute Lymphoblastic Leukaemia Diagnosis from Peripheral Blood Smear Images

We present MorphX (morphology-informed
explainable AI), an ante-hoc deep learning framework for
interpretable diagnosis of acute lymphoblastic leukaemia (ALL) from peripheral blood smear images. MorphX combines a
morphology-aware feature extraction backbone with a World
Health Organization (WHO)-aligned concept bottleneck that
predicts 28 morphological descriptors, including chromatin
texture, nuclear contour irregularity, and cytoplasmic
granularity. These concept activations are mapped through a
differentiable WHO layer to a clinically weighted scalar risk score. To align training with diagnostic priorities, we used a riskweighted objective that placed a greater emphasis on clinically important descriptors. MorphX further integrates concept-level attribution, Grad-CAM++ visual overlays, concept sensitivity scoring in the WHO concept space, and sparse counterfactual WHO perturbations that identify the minimal descriptor changes required to alter the predicted class. On average, only 7.2 of the
28 features required adjustment per case (mean |Δ| = 0.361),
indicating compact and semantically constrained counterfactual explanations. Across the ablation settings, the WHO feature supervision preserved a strong classification performance while reducing the mean squared error of the concept prediction by 15.2%. The framework achieved a classification accuracy of 94.1% and an AUC of above 0.97 on the ALL-IDB2 dataset. MorphX also generates a structured clinical report artefact that combines diagnostic output, concept vectors, risk scores, visual evidence, and counterfactual summaries. These results show that MorphX improves the interpretability of ALL image analysis
while maintaining a strong predictive performance in the present dataset.