FlashDeconv estimates cell type proportions from spatial transcriptomics data (Visium, Visium HD, Stereo-seq) with linear O(N) scalability via leverage-score-weighted structure-preserving sketching and sparse spatial graph regularization.