fidopt Interactive view
Lesions and tumours, segmentation

ISPY1-Tumor-SEG-Radiomics

First post-contrast phase of the pre-treatment breast DCE MR of a woman with locally advanced breast cancer, the whole primary tumour outlined by hand and approved by a breast radiologist (ISPY1-Tumor-SEG-Radiomics, TCIA).

16 test casesbreastimage MR
Look at the images
Read before comparing (9)
  • Pool. 163 women of the I-SPY 1 TRIAL (ACRIN 6657), with stage 2 or 3 breast cancer of at least 3 cm, treated by neoadjuvant chemotherapy at several centres from May 2002 to March 2006. The analysis result keeps those whose pre-treatment DCE has at least two post-contrast phases and whose histopathology and recurrence-free survival are known.
  • Image. A unilateral sagittal fat-suppressed T1-weighted 3D gradient echo at 1.5 T. The case is the provider's first post-contrast phase, copied as is: bias-corrected (N3), resampled to 1 mm and z-scored over the patient's three phases with CaPTk 1.8.1, the grid the outline lies on. The tumour's median z-score is 3.0 to 29.0 on the test cases, the whole image's -0.3 to -1.3.
  • One image. A method gets only this phase, the one the outline was drawn on and where the tumour enhances. The pre-contrast and second post-contrast phases are not given, nor the phase on its acquired grid, which the outline does not lie on.
  • Truth. The manual structural tumour volume (masks_stv_manual): the whole primary tumour, enhancing and non-enhancing, satellite lesions left out. A breast radiologist located it from the provider's functional tumour volume and delineated it in ITK-SNAP; a fellowship-trained breast radiologist reviewed, edited and approved every outline. It meets the main leaderboard's bar, but with no second reader its own error is unknown.
  • Pieces. Five of the ten test outlines drawn first carry 1 to 41 small pieces apart from the main body, at most 1.3% of their voxels, such as 41 around ISPY1_1145's 198 ml body; they count in the Dice and the HD95. The tumours measure 1.9 to 200 ml on the test cases. The surface Dice is at 1 mm, on the 1 mm grid.
  • Room. The provider's 3D residual U-Net reached a median Dice of 0.74 over the held-out folds of its nested cross-validation. Its shipped outlines (masks_stv_resunet, not used) score 0.65 on the ten test cases drawn first (0.01 to 0.91; 0.01 on ISPY1_1062, a dense breast enhancing strongly around the tumour); the paper does not say whether they come from held-out folds.
  • Checks. TotalSegmentator makes no breast tumour, so the labels were checked by eye: on the 20 cases drawn first, and ISPY1_1094 and 1158, every outline lies on an enhancing mass. The bright artefacts the paper reports outside the breast, on ISPY1_1183 (development) and ISPY1_1187, lie away from the outline.
  • Overlap. No other dataset of the corpus holds I-SPY 1 images. BreastNACT, UCSF's pilot study, enrolled from 1995 to 2002, and neither collection names a patient of the other; ISPY2_Syn is a later trial, 2010 to 2016. MAMA-MIA (screening segmentation 119) annotates I-SPY 1 patients again: a build of it should leave out the 30 patients used here.
  • Outside numbers. A network trained on this analysis result or on MAMA-MIA may have seen the test cases.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.

About the data

The first test case of ISPY1-Tumor-SEG-Radiomics
The first test caseClick to look at every case.
Test cases
16 test patients drawn with seed 2026 by scanner maker among the 149 of the analysis result outside development (11 GE, 4 Siemens, 1 Philips), and 14 development patients, the last of the same seeded order, each maker in proportion (10 GE, 3 Siemens, 1 Philips); ISPY1_1237 is kept out of development, its outline lying off its image's grid. One examination a patient, the pre-treatment DCE.
Truth
contours by experts the analysis result's structural tumour volume: the whole primary tumour, its enhancing and non-enhancing parts, delineated by hand in ITK-SNAP on the first post-contrast phase by a board-certified breast radiologist, from the location the provider's tumour box and functional tumour volume gave, then reviewed, edited and approved by a board-certified, fellowship-trained breast radiologist; satellite lesions left out
Licence
CC BY 3.0
14 cases for tuning, never tested

ISPY1_1161, ISPY1_1183, ISPY1_1177, ISPY1_1146, ISPY1_1189, ISPY1_1181, ISPY1_1037, ISPY1_1125, ISPY1_1132, ISPY1_1032, ISPY1_1090, ISPY1_1196, ISPY1_1222, ISPY1_1107