In this research, we propose Genetic Algorithm (GA) based clustering and classification feature selection methods for acute lymphoblastic leukemia (ALL) diagnosis. Specifically, three feature optimization algorithms, i.e. (1) filter-based enhanced Fuzzy C-Means (EFCM) + GA, (2) filter-based linear discriminant analysis (LDA) + GA and (3) embedded-based Support Vector Machine (SVM) + GA, are proposed to select the most discriminating features for robust and reliable ALL detection. In particular, to overcome limitations of original FCM with purely the intra-cluster variance measurement, the EFCM model is proposed where both inter- and intra-cluster variations are taken into account to inform GA-based discriminative feature selection. A total of four commonly used classification methods, i.e. Naive Bayes (NB), K-Nearest Neighbours (KNN), Multilayer Perceptron (MLP) and SVM, are subsequently used for ALL diagnosis with the most significant features identified by the above feature selection methods as inputs. Evaluated using ALL-IDB2, the proposed EFCM + GA feature selection approach illustrates great superiority over other feature selection and existing feature projection methods for ALL diagnosis.