Flaw detection problems in ultrasonic NDE can be considered as two-class classification problems, i.e., determining whether a flaw is present or not present. To be practical, a flaw classification method must be able to handle the uncertainties associated with interference from grain noise which leads to poor signal-to-noise ratios (SNR). In this work, the use of neural network models and statistical correlation is demonstrated for one such detection/classification problem. In particular, based on simulation studies, we wish to establish practical strategies in detecting weak volumetric flaw signals corrupted by high grain noise. An example of this type that is of recent interest is the detection of “hard-alpha” inclusions in aircraft titanium components [1]. Both the feasibility and reliability of using these classifiers are assessed. This effort was carried out in parallel with another study [2] where more traditional signal processing approaches were taken.