Efficient human-in-loop deep learning model training with iterative refinement and statistical result validation