Potential report–image association issue in CancerVerse metadata
Dear CancerVerse Team,
We understand from the dataset card that CancerVerse is longitudinal and that multiple CT volumes may belong to the same patient across different timepoints or to the same examination across different contrast phases. According to the documentation, a stable Patient ID should link scans from the same patient over time, while Encrypted Accession Number identifies the corresponding study or visit.
During our audit of CancerVerse_dataset_metadata.csv, we found that some exact, highly detailed radiology reports are shared by records with different Patient ID, different Encrypted Accession Number, and different CancerVerse ID. These cases generally have the same examination date, sex, age, and often the same contrast phase. Therefore, they do not appear to represent conventional longitudinal timepoints, which would normally share the same stable Patient ID and occur on different dates.
For example:
CV_00008426 and CV_00011852 have different Patient IDs and accession numbers but share the same examination date, sex, age, contrast phase, and an identical report describing a 22 × 28 × 24 mm right upper-pole renal lesion and two pancreatic-head IPMNs.
CV_00014767, CV_00017459, and CV_00021829 have three different Patient IDs and accession numbers but share the same examination date and an identical detailed report describing metastatic lung adenocarcinoma, confluent liver metastases, a progressive left lower-lobe mass, and adrenal metastases.
CV_00017216 and CV_00022048 have different Patient IDs and accession numbers but share the same examination date and an identical report describing small-bowel perforation, suspected bowel ischemia, ascites, and metastatic cholangiocarcinoma.
Using exact matching of the original report field, we identified 562 such report groups spanning more than one Patient ID, involving 1,230 metadata rows. Among these groups, 558 share the same examination date, sex, and age. This pattern may indicate that some volumes from the same underlying patient or examination received inconsistent de-identified identifiers, or that reports were duplicated during metadata construction.
Could you please clarify:
- Whether these records are intentionally different volumes or phases from the same underlying examination;
- Whether volumes belonging to the same patient may sometimes have different Patient ID or accession identifiers;
- Which fields should be used to reconstruct study-level and longitudinal patient-level groupings reliably;
- Whether an updated or corrected metadata file is available.
We would be happy to provide the complete list of affected CancerVerse IDs and duplicated report groups.
Best regards,
Lin
Hi @wulilin ,
Thank you for the careful check and for the clear examples. This kind of feedback really helps us improve CancerVerse, and we appreciate it.
We looked into the cases you found. Here is what we see: in our original records collected from the source, these CT volumes already had different Patient IDs and different Accession Numbers, even though they share the same report and the same examination date. So the different IDs were already in the source data we received, and we kept them when we released the dataset. This looks like an ID problem on the hospital side, not a mistake we made when building the metadata. We will contact the hospitals to check this and fix it where we can.
Here are direct answers to your four questions:
- Whether these records are intentionally different volumes or phases from the same underlying examination
It is normal to see several CT volumes with the same examination date, sex, and age when they also have the same Patient ID and Accession Number — these are real multi-phase CT volumes from the same patient, and they are correct. The cases you found do not belong to this situation; they come from the source ID problem above.
- Whether volumes belonging to the same patient may sometimes have different Patient ID or Accession Number
This should not happen by design. When it does, it comes from the source ID problem above.
- Which fields should be used to reconstruct study-level and patient-level groupings reliably
Use Patient ID to group by patient, and Accession Number to group by study.
- Whether an updated or corrected metadata file is available
We will update the metadata once the source ID problem is solved. Stay tuned!
It would help a lot if you could send the full list of affected CancerVerse IDs and duplicated report groups to wli131@jh.edu. Thank you again for helping us make CancerVerse better!
Best,
Wenxuan
Hi Wenxuan,
Thank you very much for the detailed clarification and for looking into these cases.
We have compiled the full list of affected CancerVerse IDs and exact duplicated-report groups according to your guidance. We used Patient ID for patient-level grouping and Accession Number for study-level grouping, and excluded records sharing the same Patient ID and Accession Number because they may represent valid multi-phase volumes.
The resulting list contains 562 duplicated-report groups and 1,230 affected volumes. I will send the complete workbook to wli131@jh.edu. We also retained some short and generic duplicated reports for completeness, although some of them may be legitimate independent repetitions rather than ID errors; these can be reviewed separately using the report-length fields in the workbook.
Thank you again for your help and for planning a metadata update. We also look forward to gaining access to the complete CancerVerse dataset as soon as it becomes available.
Best,
Lin