Harmful algal blooms (HABs) can have significant negative economic, environmental and health impacts. Therefore the real-time monitoring of phytoplankton is becoming increasingly critical for proper management of water bodies. This work demonstrates that instance segmentation on phytoplankton is possible on a small dataset with a large number of classes. This is accomplished in three main steps. First, 596 images of 21 different unialgal cultures were captured using brightfield microscopy. Second, these raw images were processed using traditional computer vision methods to rapidly create a binary mask. Finally, these raw images and binary masks were used to train a deep learning instance segmentation model. Experimental results show that high instance segmentation performance can be achieved for certain algae types and mixed performance for others by finetuning a Mask R-CNN deep convolutional neural network with a small but highly diverse dataset of different phytoplankton. These results show promising progress towards building a real-time on-site monitoring phytoplankton system.