fidopt Interactive view
Brain metastasis on post-contrast MR, the resected lesion only, segmentation

Brain-Mets-Lung-MRI-Path-Segs, the resected metastasis

Pre-operative post-contrast T1 MR of a lung-cancer patient with one brain metastasis at Yale, the metastasis later resected outlined by hand and scored within 20 mm of it (Brain-Mets-Lung-MRI-Path-Segs, TCIA).

10 test casesbrainimage MR
Look at the images
Read before comparing (8)
  • Pool. 111 resections of 103 lung-cancer patients at Yale; 63 had 2 to 35 brain metastases at diagnosis. The pool keeps a resection with one metastasis at diagnosis (48), of a patient with one resection (37), resected at the age of diagnosis or a year later (33; the sheet gives ages in years), with a post-contrast T1 (31). Held out: YG_JCGGWVPZ3S0Z, 26% of whose mask lies where brain extraction blanked the image, and seven possible BrainMetsMR patients (Overlap).
  • Images. The post-contrast T1 as the hospital's PACS exported it, on its own grid, acquired at Yale on every test case and at an outside facility on 4 of the 23 (3 development cases): slices 0.9 to 1 mm apart on 8 of the 23 cases (4 test cases), 5 to 7 mm apart on the others, 0.34 to 1 mm in plane, Siemens 1.5 and 3 T scanners on 21. HD-BET blanked everything outside the brain; the values are the scanner's. The FLAIR lies on another grid and is not given.
  • Truth. The core tumour, the enhancing lesion with any necrosis inside it, drawn by hand in a research PACS once the medical record had located the lesion resected, and verified by a board-certified neuroradiologist. One reader and one check: graded expert. Every mask is solid and in one piece, but development case YG_FIUSCJCISOT2's has a second piece of 0.35 ml.
  • Scored region. Only the resected metastasis is drawn, and one metastasis at diagnosis does not rule out another on the pre-operative scan. Each case's ignore region, never given to methods, takes every voxel farther than 20 mm from the drawn lesion out of the truth and the prediction before scoring, so a lesion found elsewhere counts neither for nor against a method. 20 mm is the radius of the median lesion.
  • Lesions. 3.1 to 74 ml over the 23 cases and 16 to 56 ml over the test cases (median 30 ml); every lesion lies on the side the clinical sheet names. On test case YG_JEBDFNTNS3Z1 brain extraction cut into the lesion: 12% of its mask lies where the image is zero.
  • Start. Inside the scored region, labelling the whole brain (the image's non-zero voxels) scores a mean Dice of 0.27 over the 10 test cases (0.17 to 0.38; 0.04 over the whole image), Otsu's bright voxels 0.32 (0.08 to 0.67). The surface Dice is at 1 mm, the corpus's default; six test cases have 5 to 6.5 mm slices.
  • Exposure. The masks are public on TCIA, which titles their row 'BraTS 2023 Challenge': a network trained on BraTS-METS data, which holds Yale patients, may have seen these.
  • Overlap. BrainMetsMR (Pretreat-MetsToBrain-Masks) is the same hospital's cohort, without a crosswalk. A pool case is held out when sex, age at diagnosis within a year, lung primary (small-cell or not, which the two sheets may code apart) and pack-years within a quarter, or missing, match one of BrainMetsMR's 17 lung-cancer test or development patients: seven are. The rule pairs BraTS-MET-00298 with YG_0CBM148C1MFN alone of the 111 resections, and their images agree after a rigid registration (correlation 0.77, the next patient 0.51). Registered the same way to each of the 17, the 23 pool images correlate at most 0.56 (median best 0.47). No other descriptor names the collection.

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

Test cases
10 test resections drawn with seed 2026 among the 23 of the pool, then 10 development resections, the last of the same seeded order; one resection a patient, all from Yale's brain metastasis biorepository, so no strata. The notes say how the pool was chosen.
Truth
contours by experts the provider's core tumour, the contrast-enhancing lesion with any necrosis inside it, drawn by hand on the post-contrast T1 of the metastasis later resected and verified by a board-certified neuroradiologist; no other metastasis is drawn, so a case is scored within 20 mm of this one
Licence
CC BY 4.0
10 cases for tuning, never tested

YG_3OAF908JG3XG, YG_PJ6VL4EIAVKN, YG_2VWCV5YWB078, YG_748ADE23A96L, YG_XY6SP25JK2UJ, YG_F2VHX1DGAYIC, YG_S77I03P4O3LA, YG_3YJ63A56N6VQ, YG_Q70O5ZGBYWBC, YG_FIUSCJCISOT2