MedLite XAI A Metadata Aware Calibrated and Explainable Lightweight Framework for Multiclass Skin Lesion Classification

Automated dermoscopic classification can support skin lesion assessment while reliable research requires leakage safe partitions class sensitive evaluation calibrated confidence interpretable predictions and practical computational cost; therefore this work presents MedLite XAI which combines a MobileNetV3 Small image encoder with patient metadata for seven class classification on HAM10000; a deterministic lesion grouped split allocated 7012 training images 1502 validation images and 1501 locked test images while model selection used validation macro F1 only; the selected metadata aware cross entropy model obtained 74.65 percent validation macro F1 and achieved 84.08 percent accuracy 71.59 percent macro F1 and 71.34 percent balanced accuracy on the final test set; temperature scaling reduced expected calibration error from 6.69 percent to 2.09 percent while preserving class predictions and Grad CAM analysis across 20 validation cases produced mean horizontal flip stability of 0.969; consequently MedLite XAI provides a reproducible lightweight research framework although external and prospective clinical validation remain necessary.