fidopt Interactive view
Chest CT, segmentation

PleThora

Pretreatment chest CT of a patient with non-small cell lung cancer and a pleural effusion, treated at MAASTRO Clinic (TCIA's NSCLC-Radiomics), with the left and right thoracic cavities and the effusion segmented (PleThora).

10 test caseslungimage CTresearch only
Look at the images
Read before comparing (8)
  • Cavities. A 3D U-Net outlined the lungs, a fourth-year medical student expanded each outline to the whole hemithorax, tumour, collapsed lung and hilar vessels and bronchi included, and one physician vetted and corrected it. The network's and the student's outlines agree at a Dice of 0.88 to 0.98 on the test cases (median 0.94): most of the outline is the network's. Graded semi-automatic.
  • Effusion. Drawn de novo by the student, the tumour left out, then corrected by two radiologists; the map used is the primary reviewer's, re-reviewed and corrected by another physician for version 2. The reviewers' maps agree at a Dice of 0.55 to 0.99 on the test cases (median 0.97), differing most on thin effusions, with no rule on trace fluid. The test effusions hold 10.7 to 127.9 cm3 (median 58).
  • Labels. Labels 1 and 2 are PleThora's left and right thoracic cavities, with the effusion written over them as 3, pleural_effusion as TotalSegmentator names it. On the test cases 11% to 69% of the effusion lies outside the cavities, 96% to 100% of it within 6 mm, along the chest wall. The cavities hold the tumour, collapsed lung and the hilar vessels and bronchi: they are not TotalSegmentator's lungs.
  • Orientation. TCIA warns that the masks use RPI where the DICOM uses RAI, so some tools draw them wrongly; prepare.py places them by their headers and checks that label 1 lies on the patient's left. The provider's table puts LUNG1-001's effusion on the right, but its map lies in the left cavity, on fluid: the corpus uses the maps.
  • Slices. Every PleThora CT has 3 mm slices, so along the patient's axis a surface is known to 3 mm, and the HD95 and the surface Dice at 1 mm move by whole slices. Consecutive slices of a cavity sometimes hold the same outline, up to 17 of 79 slice pairs on a test case.
  • TotalSegmentator. Version 2.18 on test case LUNG1-277, its lobes merged by side: Dice 0.954 for the left cavity and 0.939 for the right. What it leaves out is soft tissue, and it makes no effusion, which then scores 0.
  • Room. The provider's 3D U-Net reaches a cavity Dice of 0.949 and 0.956 on its 86 test cases, where two physicians agree at 0.978 and 0.983. Nine of the ten test cases here are in that model's training split, so a model trained on PleThora has seen them. No outside number compares on these cases; the effusion leaves the most room.
  • Overlap. No other dataset of the corpus holds NSCLC-Radiomics CTs. MSD_Lung's come from NSCLC Radiogenomics, at Stanford. LCTSC's institution 1 scans come from the same scanner models, but a middle slice of each of LCTSC's 60 CTs differs from every slice of the 20 CTs here.
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBImages
1KonfAI segmentation app · TotalSegmentator (pleural_pericard_effusion), the authors' packageGPUapp1 of 3 structures0.374 ±0.1720.308 ±0.13148.31 ±13.834513.92.4

The order averages each method's ranks on Dice, Surface Dice, HD95 mm, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 cases. Two neighbours are tied when a paired Wilcoxon test on the shared cases does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Workstation.

Dice: how much the organ contours overlap, 1 is perfect. Surface Dice: the share of the contours' surfaces within a tolerance of each other, 1 is perfect. HD95: distance between the contours' surfaces.

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 CTs drawn with seed 2026 among the 68 outside development, and 10 development CTs, one CT a patient. The pool is the 78 patients whose pleural effusion PleThora segmented; its 324 others, recorded without effusion by a judgement no physician is said to have checked, are left out.
Truth
contours by a program, verified by people 1 and 2 PleThora's left and right thoracic cavities (lung, tumour, collapsed lung, hilar vessels and bronchi), a network's lung outline expanded by hand by a medical student and vetted by a radiologist or a radiation oncologist; 3 the pleural effusion, drawn by hand by the student and corrected by two radiologists, written over the cavities, so that thoracic_cavity_left and thoracic_cavity_right are the cavities without the effusion
Licence
CC BY-NC 3.0 (the CTs, TCIA's NSCLC-Radiomics) and CC BY 3.0 (PleThora's masks)
10 cases for tuning, never tested

LUNG1-252, LUNG1-170, LUNG1-065, LUNG1-008, LUNG1-133, LUNG1-367, LUNG1-104, LUNG1-053, LUNG1-296, LUNG1-001