We present a texture analysis methodology that combines uncommitted machine-learning techniques and sparse feature transformation methods in a fully automatic framework. We compare the performances of a partial least squares (PLS) forward feature selection strategy to a hard threshold sparse PLS algorithm and a sparse linear discriminant model. The texture analysis framework was applied to diagnosis of knee osteoarthritis (OA) and prognosis of cartilage loss. For this investigation, a generic texture feature bank was extracted from magnetic resonance images of tibial knee bone. The features were used as input to the sparse algorithms, which dened the best features to retain in the model. To cope with the limited number of samples, the data was evaluated using 10 fold cross validation (CV). The diagnosis evaluation using sparse PLS reached a generalization area-under-the-ROC curve (AUC) of 0.93 and the prognosis had AUC of 0.70, both superior to established cartilage based markers known to relate to OA diagnosis and prognosis.