fidopt Interactive view
Lesions and tumours, segmentation

LIDC-IDRI

Chest CT of one LIDC-IDRI patient with the lung nodules of 3 mm or more that at least two of its four thoracic radiologists outlined, read from the DICOM SEGs of DICOM-LIDC-IDRI-Nodules (TCIA).

30 test caseslungimage CT
Look at the images
Read before comparing (10)
  • Readers. Four experienced thoracic radiologists read each CT, first blind, each marking nodules of 3 mm or more, nodules under 3 mm and non-nodules, then each reviewing their own marks beside the others'. Nothing forced a consensus. Only nodules of 3 mm or more are outlined, on every slice they show on; the others are points.
  • Outlines. DICOM-LIDC-IDRI-Nodules' SEGs, made with pylidc and dcmqi from the XML outlines, leave out the outline pixels, which run just outside the nodule. prepare.py reads the XML only to tell which reader drew an outline.
  • Pool. Of the 1,018 CTs, left out are the 16 of eight patients with two CTs, 7 whose series has more than one XML file, 2 with three or eight reading sessions instead of four, and 135 with no SEG. 780 of the 858 left hold a nodule two readers or more outlined, 22 of them SAROSBodyCT's.
  • Truth. Label 1 where at least two of the four readers put a nodule of 3 mm or more; the ten test truths drawn first hold 1 to 7 nodules a CT, of 20 to 7,415 mm³. Nodules under 3 mm, non-nodules and nodules one reader alone outlined are not in it: a method that outlines one is charged with its voxels.
  • Truth spread. Against the voxels at least two of the other three hold, a reader scores a median Dice of 0.76 (0 to 0.90) over the 40 reads of the ten test CTs drawn first, 0.74 (0.40 to 0.96) on the development CTs. The zeros are on LIDC-IDRI-0790, where no voxel is held by three readers, and LIDC-IDRI-0844, where one reader outlined nothing.
  • Scores. A method should not be expected to agree with the truth better than one radiologist agrees with the other three. A case's Dice pools its nodules, so the largest weighs most; HD95 mostly counts nodules found or missed.
  • TotalSegmentator. Its lung_nodules task scores a Dice of 0.70 on LIDC-IDRI-0580 and 0.42 on LIDC-IDRI-0334, four of seven nodules found, and finds none on LIDC-IDRI-0790, whose 7,415 mm³ nodule lies against the chest wall. It was trained partly on LIDC-IDRI and may have seen these CTs.
  • Images. The DICOM series in Hounsfield units, as delivered: on the ten test CTs drawn first, GE, Toshiba and Siemens scanners, pixels of 0.58 to 0.90 mm, slices 0.625 to 3 mm apart. Development case LIDC-IDRI-0850 is of a prone patient, which its header says. No mask is given.
  • Seen. LIDC-IDRI has been public with its outlines since 2011 and is in many training sets, LUNA16's 888 CTs among them: a model trained on it has probably seen these test cases.
  • Overlap. No other dataset of the corpus holds a CT of these patients: checked against AeroPath, LNDbNodules, MSD_Lung, LUNA25, RIDERLungCT and NLSTNodules, and the sixty names against every descriptor and prepare.py. ATM22 leaves its LIDC-IDRI CTs out. A later build of LUNA16, QIN-LungCT-Seg, Pulmonary-Nodules-Segmentation or SAROS must leave these sixty out.

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 LIDC-IDRI
The first test caseClick to look at every case.
Test cases
30 test CTs drawn with seed 2026 among the 727 outside development, then 30 development CTs, the last of the same seeded order. The pool, 757 CTs, is LIDC-IDRI's CTs whose readers share a nodule of 3 mm or more, one a patient, once SAROSBodyCT's 24 LIDC-IDRI patients and LIDC-IDRI-0909, whose two readers share no voxel, are left out. The provider publishes no test set.
Truth
contours by experts the voxels that at least two of the CT's four readers put inside a nodule of 3 mm or more (the 50% agreement), each reader a thoracic radiologist who outlined every such nodule slice by slice, first blind to the others, then with the others' marks in view, without forced consensus
Licence
CC BY 3.0 (LIDC-IDRI's CT images and DICOM-LIDC-IDRI-Nodules' segmentations, on their TCIA pages)
30 cases for tuning, never tested

LIDC-IDRI-0338, LIDC-IDRI-0850, LIDC-IDRI-0800, LIDC-IDRI-0209, LIDC-IDRI-0427, LIDC-IDRI-0980, LIDC-IDRI-0550, LIDC-IDRI-0815, LIDC-IDRI-0704, LIDC-IDRI-0733, LIDC-IDRI-0111, LIDC-IDRI-0722, LIDC-IDRI-0756, LIDC-IDRI-0732, LIDC-IDRI-0602, LIDC-IDRI-0016, LIDC-IDRI-0727, LIDC-IDRI-0692, LIDC-IDRI-0091, LIDC-IDRI-0128, LIDC-IDRI-0785, LIDC-IDRI-0966, LIDC-IDRI-0254, LIDC-IDRI-0582, LIDC-IDRI-0939, LIDC-IDRI-0633, LIDC-IDRI-0258, LIDC-IDRI-0480, LIDC-IDRI-0529, LIDC-IDRI-0969