Fig. 1

Algorithm construction and training schematic: A The algorithm training phase – Slides from normal colons were selected and ganglion cells were manually annotated The algorithm was then trained on these annotated fields. 10% of the data was reserved for further analysis. B The algorithm analytical performance phase—included reserved data, as well as additional slides from cases with clinical suspicion of HSCR. The algorithm was run on un-annotated data to produce annotations of its own, for which a pathologist has provided feedback. C The algorithm was then run on an additional 50 cases with clinical suspicion of HSCR and provided image sets of the best ganglion cell candidates it could find along with a their respective scores (0 to 1). The pathologist reviewed the image sets and provided his own score (1 to 5). The overall HSCR status of a given case was determined through a combination of the algorithm and pathologist scores (according to previously empirically determined decision criteria)