fidopt Interactive view
Lesions and tumours, segmentation

MSD Lung

Thin-section chest CT of a patient with non-small cell lung cancer before surgery, from Stanford University, with the tumour segmented (Medical Segmentation Decathlon Task06, from TCIA's NSCLC Radiogenomics).

10 test caseslungimage CT
Look at the images
Read before comparing (8)
  • Truth. On each of the 20 cases built the decathlon's mask is, voxel for voxel, NSCLC Radiogenomics' segmentation of the patient: drafted by an algorithm, edited by a thoracic radiologist with more than five years of experience and reviewed by a second (Bakr et al. 2018). The data paper says instead that one radiologist marked a representative cross-section; the masks are 3D.
  • Left out. lung_022's mask is its patient's segmentation mirrored head to feet and moved 8 slices, a Dice of 0.52 with it. It had been drawn as a test case; lung_051 took its place. The 42 training cases not built were not compared with the collection.
  • Tumours. Small: 0.8 to 12 ml on the test cases, median 2.3 ml, a mask in up to 12 pieces, many part solid or ground glass. On a small tumour a voxel more or less moves the Dice.
  • Images. The decathlon's NIfTI files in Hounsfield units, clipped at -1024, transposed without resampling to the closest RAS orientation. On 10 of the 20 cases the headers give 1.2445 to 1.2457 mm between slices that lie 1.25 mm apart, so distances along the body read up to 0.4% short, for truth and method alike. lung_018 holds metal up to 20349 at the front of the chest.
  • Room. On the challenge's hidden test set the 19 entries' mean Dice ran from 0.00 to 0.69, median 0.51, nnU-Net 0.69. The surface Dice is at 2 mm, the error the challenge's clinicians would accept.
  • Outside numbers. None compare: the challenge ranks on its hidden test set. The test cases are public training cases of the decathlon, so a network trained on its lung task has seen them.
  • Checks. On the 20 cases built every label map holds 0 and 1 on its image's grid; on test case lung_004 TotalSegmentator puts 99% of the tumour within about 5 voxels of its left lung lobes, on the patient's left.
  • Overlap. NSCLCRadiogenomics_CTPET_Syn holds PET/CT sessions of this collection: its R01 pool starts at R01-082, after the patients of the 20 cases here (R01-004 to R01-081).

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 MSD Lung
The first test caseClick to look at every case.
Test cases
10 test cases drawn with seed 2026 among the 52 training cases outside development, and 10 development cases. lung_022, whose label is misplaced, is left out. The labels of the decathlon's 32 test scans were never released.
Truth
contours by experts the lung tumour (label 1): on every case built the decathlon's mask is, voxel for voxel, NSCLC Radiogenomics' segmentation of the patient, drafted by an algorithm, edited by a thoracic radiologist and reviewed by a second
Licence
CC BY-SA 4.0
10 cases for tuning, never tested

lung_033, lung_045, lung_044, lung_014, lung_054, lung_009, lung_026, lung_057, lung_069, lung_023