get_largest_connected_component_mask materializes redundant nonzero index arrays before bincount #8974

Description

get_largest_connected_component_mask (monai/transforms/utils.py) ranks component sizes with:

nonzeros = features[lib.nonzero(features)]
features_to_keep = lib.argsort(lib.bincount(nonzeros))[::-1][:num_components]

lib.nonzero(features) materializes one full index array per spatial dim, and features[...] gathers a full copy of every non-background voxel, just to drop background before bincount. That is unnecessary: bincount over the whole field already counts every label, and background is index 0.

Proposed fix

counts = lib.bincount(features.reshape(-1))
counts[0] = 0  # ignore background
features_to_keep = lib.argsort(counts)[::-1][:num_components]

Bit-identical output (verified), without the nonzero index arrays or the gathered copy, so faster and lower peak memory.