The basic idea is that you do some steps that trigger a document leak, then get a heap dump. Typically this involves loading the test page, then navigating away, clearing caches, triggering a GC and capturing a snapshot. The snapshot is dumped as JSON file in /tmp; you can then drag this snapshot into the view which lives in Tools/GCHeapInspector, and see why objects are referencing the leaked document.
Let me know if you have questions. Improvements to the inspector page are welcomed.