What is required in HumanLoopInput.InputContent for start_human_loop

I have built a custom key-value extraction workflow that leverages textract Tables and Forms. It does a whole heap of post processing using the output of Textract to extract a small number of highly important fields from documents that are of very poor quality.

My client would like a human-in-the-loop to make minor changes to the results where certain fields are missing. I think that the sagemaker_a2i_runtime.start_human_loop is the perfect tool for this.

I want to send the human reviewer, the input image and the current Key-Value pairs that I have extracted and have them find any that are missing, or mark them as not there. I have setup and tested the textract.analyse_image workflow with a human reviewer and like the results.

What structure and data fields do I need to set in the HumanLoopInput field of sagemaker_a2i_runtime.start_human_loop to get this to work. I assume that it will look something like a dictionary of current K-V pairs and the s3 image file but I cannot find any documentation on how to do this.