Protein production has experienced great advances in recent years. In particular, high throughput protein production, coupled with the use of robotics, outputs thousands of mixtures in micro-array wells. To detect the presence of protein crystal formation, images of these wells are acquired regularly using robotic cameras. Traditionally, a crystallographer would manually process each image--identifying the wells that resulted in protein crystal formation. This manual inspection process is slow and given the high rate of mixture output, it has become near impossible for crystallographers keep up. Our aim is to create an automated method of detecting which wells have crystals and which ones do not. We make use of a neural network that is trained based on manually classified ground truth data. After it is trained, the automatic classifier would give a binary output --a value of one for the detection of crystals and precipitates in images and a value of zero otherwise. In our previous papers, the core methods of using multi-scale Laplacian image representation to extract image features and the implementation of the neural network classifier were discussed. Here we present a new, optimized approach to training the neural network and results from a large-scale test. We claim that the neural network can be better trained if the training image dataset is optimized in the sense that ambiguous images are removed during the initial training processes. Incremental training is implemented so that the network can be improved as more data becomes available. From initial results with training based on a 6,000 optimized image dataset, the accuracy of the improved classifier approaches 95% in identifying a wide array of images.