Datasets:

ArXiv:
License:

Mismatch Metadata.JSON and SkinB2

#10
by anrooijb - opened

Dear Histai Team,

My sincere gratitude for providing this large and insightful dataset and all the work that entails.
I feel that there might be an issue with the correlation between the metadata JSON case files matching case numbers as uploaded to skin-B2, although the problem could be on my end.

For instance, histai/HISTAI-skin-b2/case_00193 describes five fragments of a lipoma, but the associated case at https://huggingface.co/datasets/histai/HISTAI-skin-b2/tree/main/case_00193 to my best interpretation shows only 4 fragments of what is not a lipoma.

Similarly, histai/HISTAI-skin-b2/case_00026 metadata describes an epidermoid cyst, but the associated case at https://huggingface.co/datasets/histai/HISTAI-skin-b2/tree/main/case_00026 to my best interpretation displays a quite beautiful basal cell carcinoma.

Is there an issue with the metadata of Skin-B2 or how I converted it to a excel sheet or does the numbering of the cases at huggingface not match the case numbering in the metadata?

My appreciation in advance for the help,

Thank you very much for all your efforts Histai group.

We see the same problem in HISTAI-skin-b2. It looks like the report text is attached to the wrong case for part of
skin-b2, while the ICD-10 code in the same metadata record still matches the slide.

Slides. I'm a dermatopathologist and read two skin-b2 slides blind, before opening the reports:

case report conclusion icd10 slide
case_15775 (slide_H&E_0) Kaposi's sarcoma D22.5 melanocytic nevus
case_15017 (slide_H&E_0) Fibropapilloma of the skin of the left breast areola D22.9 melanocytic lesion

The two cases above fit the same pattern:

  • case_00026: text epidermoid cyst, code C43-C44, slide BCC.
  • case_00193: text lipoma, code C44, slide not a lipoma.

In every case named so far the text is wrong for the slide, and where the lesion was named, the code is right. Our
copies of the two files are byte-identical to main (their sha256 reproduces the blob ids).

Metadata-wide. I counted only clear-cut records:

  • Code C43/C44 vs D22/D23/L72/L82.
  • A conclusion naming a carcinoma, melanoma or sarcoma, vs one naming only a nevus, keratosis, papilloma, cyst,
    lipoma or fibroma, with no negation or differential.

The conclusion contradicts the same record's icd10 on malignant vs benign in 1,090 of 14,587 comparable skin-b2
cases (7.5%)
, and in 0 of 1,083 comparable skin-b1 cases. A random sample of 25 all read as plain
contradictions, for example:

  • case_10111, D22.5: "Basal cell carcinoma of the skin of the neck. Excised with clear margins. pT1"
  • case_18015, D22.3: "Desmoplastic melanoma of the skin of the shoulder area. pT4b."
  • case_12466, D22.5: "Basal cell carcinoma of the skin. Resection margin involves the tumor."
  • case_04437, C43.2: "Compound melanocytic nevus of the trunk skin."
  • case_01606, C44.5: "Intradermal melanocytic nevus of the skin of the neck."

This is a lower bound: conflicts within malignant or within benign (e.g. case_04589, BCC text with C43.3) aren't
counted. The counts are the same in the original metadata.json on this repo (revision c687f2fd74ec) and in the
Google Drive file the README now links to, so it isn't from the re-translation.

Could you check how the skin-b2 reports were linked to case IDs and publish a corrected mapping?

Thanks again!

Sign up or log in to comment