[Hendrik-code]

The defect

fetch_image_config has always returned (out_list, err), where err collects every path that is not on disk. Every caller discarded it:

# generate_augmentations.py
data_list, _ = fetch_image_config(config_data=data_config, split="TRAINING")

# train_monai.py
train_list, err_train = fetch_image_config(...)   # err_train never read
val_list,   err_val   = fetch_image_config(...)   # err_val never read

So a data config naming 500 subjects of which 400 are missing augments 100 of them, or trains on 100 of them, and prints nothing at all. Silent data loss is the failure mode that costs a whole run before anyone notices.

Second defect in the same filter: it only checked IMAGE.

if not os.path.exists(input_img_path):
    err.append([input_img_path, "path error"])

A pair whose label was missing passed the filter and failed much later, inside a process_map worker, with an opaque nibabel error.

The fix

Check both paths, and add report_missing(err, split) — a count, the first few paths, and how many more — called from both scripts.

Tests

unit_tests/test_missing_inputs.py: a pair with a missing label is rejected and named; a complete pair is accepted; the report names paths, stays silent when nothing is missing, and truncates a long list; and both scripts call it.

Verified failing before the fix and passing after. Full suite unchanged against baseline; ruff check . clean.

🤖 Generated with Claude Code