Ovarian-Bevacizumab-Response: how to get 253.8 GB of pathology slides
Ovarian cancer bevacizumab response dataset · 286 H&E whole-slide images · 78 patients · CC BY 4.0
The data comes in two parts: three small tables download directly with curl; the 286 slides, about 253.8 GB, have no ordinary download link and are distributed only via Aspera. PathTrove receives TCIA slide packages whole with ascli, the official Aspera command-line tool; nothing needs clicking in a browser, so it runs on a server. No account, approval or data use agreement is needed.
1. What to download
File
Contents
Size
How to get it
Pathology slides
286 SVS files; H&E, 20×, 0.5 µm/pixel
About 253.8 GB
Whole package via the ascli command line
new_CA125-data_20230207.xlsx
Response label per slide (effective / invalid) and CA-125 before and after treatment
26 KB
Direct download with curl
Final-patient_list.xlsx
Per patient: age, diagnosis, FIGO stage, bevacizumab regimen, and dates of surgery, recurrence and death
18 KB
Direct download with curl
Slide manifest CSV
Patient ID of each slide, and the ID used for online viewing
18 KB
Direct download with curl
The official page and the paper still say 288 slides (162 effective, 126 invalid). The current Version 2 removed 414056O.svs and 414056P.svs, leaving 286 (160 effective, 126 invalid). The label sheet name Ovary.effective-162 is also the old count.
The three files should be 26,219, 18,640 and 18,721 bytes; the slide manifest has 286 data rows.
3. Download the slides
First install ascli and the transfer engine ascp (Ruby 3.1 or later is required), then receive the whole package with the Faspex package link from the TCIA page:
Only the whole package can be received. Appending a subdirectory to the command returns a server 500 error.
--ts is required. The default transfer settings sent by Faspex set the minimum rate to 0, and the transfer stays at 0 bytes; setting the minimum to 100 kbps fixes it.
To resume after an interruption, rerun the same command. Unfinished files are kept as *.partial and *.aspera-ckpt and renamed automatically when complete; do not empty the folder and start over.
Data starts arriving about 100 seconds after launch; seeing 0 bytes during that time is normal. PathTrove measured TCIA Aspera transfers from mainland China at about 0.35–0.9 MB/s, which puts the whole package at 3–8 days. The transfer uses UDP port 33001, which must be open in the firewall. Reserve at least 260 GB of disk space.
These commands have been tested on the TCIA AML-Cytomorphology (all 97 files match the official MD5), CMB-GEC, SLN-Breast and DLBCL-Morphology packages, downloading real images. The slide package of this dataset (Faspex package 744) has not yet been tested with them; the package link is taken from the official TCIA page.
4. Matching tables to slides
slide_id in the slide manifest = the file name without .svs = Image No. in the label sheet.
patient_id in the slide manifest = Patient ID in both xlsx files. A patient can have several slides, so split data by patient.
In the label sheet, 5 file names are written differently from the slide manifest, and 4 slides have different patient IDs in the two tables. The split JSON handles these cases.
5. Check after download
cd ovarian_bev/slides
md5sum -c *.sums # if the package includes official MD5 lists, check every file
find . -iname '*.svs' | wc -l # should be 286
du -sh . # about 253.8 GB
TCIA Faspex packages usually include official MD5 lists (*.sums) in the root folder; whether this package does must be confirmed after receiving it. If not, rely on the file count, the total size and whether openslide can open each file.
6. Use in experiments
The split JSON contains labels and a patient-level split for 265 slides from 73 patients (51 train / 11 validation / 11 test patients); another 21 questionable slides are listed separately and excluded for now. The preparation package includes a check script; after unzipping, run:
The script checks each file in the list: whether it exists, whether names are duplicated, and whether the file header is SVS/TIFF. Once it passes, read images and labels by split.
