CT spleen segmentation (MSD Task09)

What is measured. Outlining the spleen on axial abdominal CT slices. Data: Medical Segmentation Decathlon Task09 Spleen (Simpson et al., arXiv:1902.09063; Antonelli et al., Nat. Commun. 2022), 41 labelled portal-venous CT volumes of patients treated for liver metastases at Memorial Sloan Kettering, with expert spleen masks. The 20 unlabelled test volumes of the archive are not used.

Split by patient. 8 of the 41 labelled volumes (seed-hash rank of the case id with seed 20261006) are held out for testing and the 33 others give the training slices, so no patient is on both sides; disjoint_by: volume re-checks it when the subset is built.

Items. Every axial slice with at least 100 spleen voxels at native resolution is cut from its volume, put in radiological display orientation (anterior up, patient right on the image left), windowed to the soft-tissue range [-125, 275] HU and rescaled to 0-255, then resized to 224 x 224 (bilinear; the mask with nearest neighbour): 1,043 slices. 100 test slices are drawn from the 218 slices of the 8 held-out patients (spleen_2, 3, 13, 17, 18, 19, 49, 61) and 500 training slices from the 825 of the others, by seed-hash rank. Each item carries its pixel spacing in mm after resizing, so HD95 is in millimetres. (exc) The benchmark is 2-D: models see one slice at a time, not the volume, so 3-D networks are evaluated slice-wise and the numbers are not comparable with the Decathlon's volumetric leaderboard.

Zero-shot (dice_zero_shot, hd95_zero_shot). A segmentation model receives the windowed slice and the target name "spleen" and returns one binary mask of the same size; nothing is trained here.

Supervised (dice_supervised, hd95_supervised). Frameworks with no pretrained weights are trained from scratch on the 500 training slices and their masks, then predict the 100 test slices. Kept apart from zero-shot.

Metrics. Dice and HD95 per slice (HD95 = 95th percentile of the pooled surface distances in both directions, face connectivity, as medpy), then the mean over slices. An empty prediction scores Dice 0 and an HD95 equal to the slice diagonal (the largest distance the image allows). 95 % bootstrap interval over test slices with whole patients resampled (cluster bootstrap on volume), because slices of one patient are correlated.

Caveats. With 8 test patients the interval is wide; it says so. Generalist medical segmenters (TotalSegmentator, MedSAM-style models) were often trained on collections that include the Decathlon spleen volumes; their zero-shot numbers may include training patients (a known overlap to report with the results).

Surface Dice and per-patient Dice (nsd_zero_shot, nsd_supervised, volume_dice_zero_shot, volume_dice_supervised; metrics review). The Decathlon ranks on Dice and the normalised surface distance (NSD): the share of the two outlines' boundary pixels lying within a tolerance of the other outline (MONAI's compute_surface_dice voxel form, the boundary of HD95). The tolerance is the Decathlon's for the spleen task, 3 mm (Antonelli et al., Nat. Commun. 2022, Methods, "Assessment method": the values were chosen per task by the clinicians segmenting each organ), measured in each slice's pixel spacing; it is computed per 2-D slice here, then the mean, so it is not the Decathlon's 3-D value. The per-patient Dice pools each patient's test slices (Σ 2|A∩B| / Σ (|A| + |B|)) and averages the 8 patients; with the bootstrap over patients, its interval is that of the 8 patient values.

Data

• msd_spleen: url https://msd-for-monai.s3-us-west-2.amazonaws.com/Task09_Spleen.tar, licence CC-BY-SA-4.0

Subset

• seed 20261006
• test: 100 item(s)
• train: 500 item(s), disjoint by volume
• results at atlas-n100-v1 use subset ffa2eae7f911d3ed

Metrics

• dice_zero_shot — Dice (zero-shot); protocol zero_shot; higher is better; range [0.0, 1.0]; 95% bootstrap CI, 1000 resamples, grouped by volume
• hd95_zero_shot — HD95 (zero-shot); protocol zero_shot; lower is better; range ≥ 0.0; 95% bootstrap CI, 1000 resamples, grouped by volume
• dice_supervised — Dice (trained from scratch); protocol supervised; higher is better; range [0.0, 1.0]; 95% bootstrap CI, 1000 resamples, grouped by volume
• hd95_supervised — HD95 (trained from scratch); protocol supervised; lower is better; range ≥ 0.0; 95% bootstrap CI, 1000 resamples, grouped by volume
• nsd_zero_shot — NSD at 3 mm (zero-shot); protocol zero_shot; higher is better; range [0.0, 1.0]; 95% bootstrap CI, 1000 resamples, grouped by volume
• volume_dice_zero_shot — Per-patient Dice (zero-shot); protocol zero_shot; higher is better; range [0.0, 1.0]; mean over volume; 95% bootstrap CI, 1000 resamples, grouped by volume
• nsd_supervised — NSD at 3 mm (trained from scratch); protocol supervised; higher is better; range [0.0, 1.0]; 95% bootstrap CI, 1000 resamples, grouped by volume
• volume_dice_supervised — Per-patient Dice (trained from scratch); protocol supervised; higher is better; range [0.0, 1.0]; mean over volume; 95% bootstrap CI, 1000 resamples, grouped by volume

Test items are 224 x 224 windowed CT slices (input) and 0/1 spleen masks (expected output). Three items of the held-out patients:

item_id	patient	slice	spleen pixels at 224²
msd-spleen_2-z078	spleen_2	78	1,567 (1.82 mm pixels)
msd-spleen_17-z074	spleen_17	74	1,000 (1.40 mm pixels)
msd-spleen_18-z123	spleen_18	123	506 (2.21 mm pixels)

A perfect mask scores Dice 1 and HD95 0 mm; predicting nothing on msd-spleen_18-z123 scores Dice 0 and an HD95 of 700 mm (the slice diagonal).