Computer texture analysis methods use texture features that are traditionally chosen from a large set of fixed features known in literature. These fixed features are often not specifically designed to the problem at hand, and as a result they may have low discriminative power, and/or may be correlated. Increasing the number of selected fixed features is statistically not a good solution in limited data environments such as medical imaging. For that reason, we developed an adaptive texture feature extraction method (ATFE) that extracts a small number of features that are tuned to the problem at hand. By using a feed-forward neural network, we ensure that even nonlinear relations are captured from the data. Using extensive, repeated synthetic ultrasonic images, we compared the performance of ATFE with the optimal feature set. We show that the ATFE method is capable of robust operation on small data sets with a performance close to that of the optimal feature set. Another experiment confirms that our ATFE is capable of capturing nonlinear relations from the dataset. We conclude that our method can improve performance in practical, limited dataset situations where an optimal fixed feature set can be hard to find.