Currently, OSU has an expert annotated database of FL images and as mentioned above the future algorithms will be trained and tested using this database. However, it will be desirable to have a semi-automatic annotation system to increase the size of the database as much as possible. In this WP, we will develop semi-automatic microscopic image annotation algorithms.

We will take advantage of the recent developments in on-line adaptive learning algorithms [33,34,35] to develop a semi-automatic system. Recently, we developed an active learning method using the well-known Least Mean Square (LMS) adaptive filtering framework and successfully applied to wildfire smoke detection [33]. The adaptive annotation system will be iterative in nature. The system will provide an initial guess to the expert and it will learn from the actions of the expert and update the annotated image whenever expert makes a correction or revision. In the LMS based active learning method weak decision rules covering an aspect of the annotation strategy are linearly combined and the weights of individual decision rules are updated according to the decisions of the expert by calculating the stochastic gradients.

In this WP, we will first try to identify subalgorithms covering an aspect of the overall decision process of the FL image annotation by a pathologist. We will then construct an adaptive scheme updating the weights of the subalgorithms in an iterative manner. We will use both the recently developed LMS based scheme and other online learning methods, such as Adaboost and the weighted majority algorithm [36].