Block reject count during index split

After performing a large purge from many of our tables, we are starting to see high waits due to index contention.  In many cases, we are seeing that index splits are consuming a high level of resources.  Doc ID 2281781.1 discusses the need to coalesce or rebuild the indexes after such large delete.  The document makes this statement "The high amount of time appears because large index block reject count before the segment is extended or a block split occurs"

I am unable to find any documentation on the internal process of an index split.  I know that a large delete will leave empty or "dead" blocks in the index pages.  What causes an empty, or near-empty block to be rejected as a candidate to be used in a split?