Dear Organizers, Before our final training runs we checked the integrity of the training release and found three things we think are unintended. Figures are attached, one per affected pair. ### 1. Five byte-identical images are published under two station names, with conflicting masks By MD5, the 524 station-named frames are only **517 distinct images**. In five of the duplicate pairs, both members are valid station frames — both have a visibility row, and the two rows are *identical* — yet the same image carries two different masks: - `center_1_case_11_6L` = `center_1_case_11_7L` - `center_1_case_3_11L` = `center_1_case_3_13R` - `center_1_case_10_6R` = `center_1_case_10_7R` - `center_1_case_15_6R` = `center_1_case_15_7R` - `center_1_case_12_6R` = `center_1_case_12_7R` Scoring one annotation against the other with the challenge metric, the five pairs average **Task 1 Dice 0.830 / normalised HD 0.121** and **Task 2 Dice 0.839 / normalised HD 0.110**. These are class-level disagreements rather than boundary jitter — whole structures are assigned to different classes, including weight-3 classes present in only one of the two annotations. The attached figures show this directly. ${image?synapseId=syn77325082} ### 2. Two `na_station_1` frames duplicate a station frame that is still in the release `center_3_case_3_na_station_1.png` is byte-identical to `center_3_case_3_12R.png`, and `center_7_case_3_na_station_1.png` to `center_7_case_3_10L.png`. Both station-named copies are still in the release and, like the extra frames, have no visibility row — so it looks as if the frame was demoted from its station but the station-named copy was never withdrawn. More generally, exactly six station-named frames have no visibility row (524 − 6 = 518), and all six are either byte-identical duplicates or otherwise anomalous: `center_3_case_3_12R`, `center_7_case_3_10L`, `center_7_case_3_12L`, `center_7_case_2_13R`, `center_7_case_2_13L`, `center_2_case_7_10L`. Should all six be reated as frames whose filename station is not to be trusted? ${image?synapseId=syn77325081} ### 3. `center_7_case_2_13R.png` — the mask looks incomplete 93.7 % background with one non-background class (*Instrument*), against a dataset median of 18.0 % and a next-highest file of 43.7 %. The same image, published as `center_7_case_2_9.png`, has a complete 15-class annotation. ${image?synapseId=syn77325088} --- ### Two questions **a) Is the segmentation ground truth a function of the image alone, or is it conditioned on the station the frame was selected for?** In those five pairs the pixels and the visibility rows are identical, yet the labels differ by station name. If this is intentional, an image-only model has an irreducible error floor and it would help to know. If not, it would be worth checking whether the same duplication exists in the test set. **b) May the station identifier in the filename be used as a model input for Tasks 1 and 2?** The rules are explicit for Task 3 and we have complied; the Task 1 / 2 case is not addressed, and the answer depends on (a). Thanks, Jmees Inc. Shunsuke Kikuchi