Lean premixed turbulent methane-air flames have been investigated using direct numerical simulations (DNS) for different Karlovitz numbers (Ka), ranging from 65 to 3350. The flames are imposed to a high intensity small-scale turbulent environment, corresponding to high Ka conditions, and the effect on the flame structure is investigated during the transition from initial laminar flame to highly distorted turbulent flame. The focus is on the internal structure of different sub-layers of these flames. The preheat layer, fuel consumption layer and oxidation layer are characterized by the distribution of formaldehyde, the fuel consumption rate and the CO consumption rate, respectively. Different measures that quantify sub-layer thickness for turbulent flames have been defined and analyzed. The flame brush is broadened with time while the local thickness (excluding large scale wrinkling) of all three layers initially show thinning due to the interaction of the flame with the turbulent flow field. As time passes, the local thickness of the preheat layer and fuel consumption layer are restored while the oxidation layer remains thinned due to suppression of CO consuming reactions. As Ka increases there is an increasing probability of finding thinned, large gradient regions in each of these sub-layers. The contribution to the evolution of flame thickness from normal strain rate, chemical reaction and normal and tangential diffusion is analyzed in terms of a gradient transport equation. The relative size of the terms changes as Ka increase and, in particular, the term due to chemical reactions loses its relative significance. The observed thinning of the local flame structure is attributed to the preferential alignment of the flame normal with the compressive strain rate eigenvectors. Such alignment provides a mechanism for the flame thinning, consistent with the behavior of non-reacting scalars. A preferred angle of about 20 degrees is observed between the flame normal and the compressive strain rate eigenvector.