The output of HoVerNet patch-based inference (80x80) does not match its input (270x270) and it cannot be corrected by resizing since it is the result of the center crop of the input (due to handling of borders). This will pose a problem when we want inference on larger ROI (>270) since using the current MONAI SlidingWindowInferer, we need to take care of each patch's borders (270-80) as well as the border of the giving ROI.

The goal here is that giving an ROI of 1000x1000 for instance return instance segmentation results with size of 1000x1000.