Yes, box-cox can be used, even if the lognormal itself isn't a perfect match for the actual distribution, with the correct choice of lambda.