Learning Sparse Interpretable Features For NAS Scoring From Liver Biopsy Images