Using artificial neural networks (NNs) in microwave medical diagnosis is recently of great interest in various problems such as early breast cancer detection, brain stroke, and leukemia monitoring. NNs facilitate the process by directly assessing the presence and properties of the tissues based on the scattered field values. Although the reported studies obtained successful results through the application of NNs to microwave diagnostic problems, they used large numbers of input data. The NN input, referred to as features, for microwave diagnosis is composed of scattered fields namely antenna transmission and reflections at the frequency of choice. Large input data increase both the number of required training samples and computational cost. Optimizing the number of antennas and frequency of operation is therefore critical to improving the performance of NN-based medical diagnosis. This work considers the correlations between the effects of different frequencies and receiver/transmitter (Rx/Tx) antennas separately in order to objectively reduce the number of features. Optimized feed-forward NNs are applied to detect the presence of object(s) with permittivity value above the predefined level within the solution domain. It is performed by designating various permittivity values to the internal object(s). Promising results were obtained by reducing the number of features approximately seven times.