fidopt Interactive view
Lungs and heart on chest X-rays, 2D segmentation

Montgomery County chest X-rays, lungs and heart

A frontal chest radiograph from Montgomery County's tuberculosis screening, 2D, its lungs drawn under a radiologist's supervision (National Library of Medicine) and its heart drawn by researchers and reviewed by a clinician (PhysioNet).

13 test caseslungimage XR2D, distances in mmresearch only
Look at the images
Read before comparing (7)
  • Images. 80 normal radiographs and 58 with tuberculosis, 0.0875 mm a pixel, kept as stored in 8-bit grey values although the ReadMe speaks of 12 bits. The set names no patient, so each radiograph is taken as its own; development cases MCUCXR_0074_0 and MCUCXR_0085_0, and MCUCXR_0083_0 and MCUCXR_0390_1, as test case MCUCXR_0367_1 and development case MCUCXR_0028_0, share sex and age but show different people.
  • Lungs. The set's own masks, drawn by hand at the National Library of Medicine under a radiologist's supervision, leaving out the lung behind the heart and inferring the outline where disease hides it. The set names masks by image side: leftMask is the patient's right lung, and prepare.py checks that the heart lies nearer the lung it calls left.
  • Heart. PhysioNet's masks (Duvieusart et al. 2023), drawn by non-medical researchers and reviewed by an experienced NHS clinician, as the pixels where radiation passes through the heart. On average 2% of the heart, 9% at most, lies on a lung, and those pixels stay lung. Nobody measured agreement between readers on either label.
  • Labels. lung_left (1), lung_right (2) and heart (3), the patient's sides: lung fields of a radiograph under CT names. Brioso et al.'s clavicles are left out, as they cross the lungs.
  • Room. On the first ten test radiographs another patient's labels, stretched to the image, score a Dice of 0.73, 0.75 and 0.67 on the left lung, right lung and heart, but a surface Dice of 0.04 at 1 mm, about eleven pixels: the Dice separates little here, the surface Dice a great deal.
  • Overlap. The Kaggle 'Chest Xray Masks and Labels' copy, Brioso et al.'s clavicles and PhysioNet's hearts hold the same radiographs: a later dataset from them derives from this one.
  • Licence. The radiographs download freely, but NLM's ReadMe asks that they not be shared outside one's research group: the corpus is built locally, and passing it on needs NLM's agreement. The hearts are under ODC-By 1.0.
AccuracyCost
#MethodDiceSurface DiceHD95 mmTime sRAM GBGPU GBImages
1TorchXRayVision PSPNetCPUapp0.814 ±0.0250.202 ±0.02727.59 ±5.09132.9

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 13 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 Laptop.

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
13 test radiographs drawn with seed 2026 among the 138, normal and tuberculosis in proportion, seven and six, then 12 development radiographs, the last of the same seeded order, those the test draw rests on kept, eight normal and four with tuberculosis. One radiograph is one case.
Truth
contours by experts the lungs drawn by hand by the set's makers at the National Library of Medicine under a radiologist's supervision, the part behind the heart left out; the heart drawn by hand by non-medical researchers on PhysioNet's 'Heart and lung segmentations for MIMIC-CXR/MIMIC-CXR-JPG and Montgomery County TB databases', each mask reviewed by an experienced NHS clinician and corrected
Licence
NLM's terms for the radiographs and lungs (no sharing outside one's research group); ODC-By 1.0 for the hearts
12 cases for tuning, never tested

MCUCXR_0028_0, MCUCXR_0085_0, MCUCXR_0063_0, MCUCXR_0083_0, MCUCXR_0074_0, MCUCXR_0006_0, MCUCXR_0029_0, MCUCXR_0313_1, MCUCXR_0282_1, MCUCXR_0311_1, MCUCXR_0390_1, MCUCXR_0046_0