Small lung nodules get a second opinion: a four-model ensemble lifts Dice by 0.01118 on external LNDb over one Mean Teacher
Synopsis
With voxel-level annotations on only 10% of CT scans, a heterogeneous ensemble of four independently trained models was applied to candidate-centred segmentation of 3–<6 mm pulmonary nodules. Against a single enhanced Mean Teacher, it raised Dice by 0.01488 on an internal held-out set and by 0.01118 on external LNDb without fine-tuning, with both confidence intervals excluding zero and concordant directions across three training seeds. The gain came with higher recall and lower precision, and did not establish that heterogeneous objectives beat homogeneous four-model ensembles.
Interpretation
Candidate-centred segmentation of small pulmonary nodules can gain a consistent Dice improvement under a small annotation budget by combining four independently trained models as a weighted probability average instead of relying on one enhanced Mean Teacher. The prior condition was a single enhanced Mean Teacher, and one model struggled to address both foreground omission and boundary expansion, two errors that pull in opposite directions for small nodules. Internal held-out Dice rose from 0.76428 to 0.77915, a paired difference of 0.01488 with a CT-cluster 95% CI of 0.00665–0.02347 and Holm-adjusted P = 0.00110; external LNDb rose from 0.52586 to 0.53705, a difference of 0.01118 with a CI of 0.00314–0.01892 and Holm-adjusted P = 0.00673.
The combination is a weighted average of the four models' output probabilities followed by binarisation at a fixed threshold, with weights and threshold fitted on the development selection set and then frozen. The four members share a three-dimensional residual U-Net backbone but are trained independently from random initialisation, and three carry recall-oriented Tversky loss, precision-oriented Tversky loss, or that plus a baseline-anchoring term. The locked weights were 0.226690, 0.308426, 0.430849 and 0.034035 with a binarisation threshold of 0.567185, fitted only on 69 nodules of 3–<6 mm in the development selection set; the internal test set held 62 nodules from 51 CT scans and external LNDb held 436 nodules from 176 CT scans, with no fine-tuning or recalibration externally.
Perspective
The results are limited to candidate-centred three-dimensional segmentation, with inputs taken as 48×48×48 voxel patches centred on nodule candidates, and do not cover candidate detection or whole-CT inference. The annotation budget is fixed at 35 of 350 development-training CT scans with voxel contours; the internal test set holds 62 nodules of 3–<6 mm from 51 CT scans, and external testing uses 436 nodules from 176 LNDb CT scans with no fine-tuning, recalibration or threshold reselection. The authors propose prospectively locking the heterogeneous and strongest homogeneous ensembles, repeating label-subset selection, and evaluating new independent cohorts.
The authors state that the results do not establish that the population-level Dice improvement exceeds 0.01, nor that heterogeneous training objectives provide an independent benefit beyond general deep-ensemble effects. Homogeneous four-model controls behaved differently across cohorts: four enhanced Mean Teachers improved internally but dropped externally, and the exploratory equal-weight four-standard-Mean-Teacher ensemble had the highest external Dice. The gain reflects a shift toward recall, with precision and some surface and volume metrics worsening. Public identifiers did not allow confirming whether any person contributed more than one CT scan, so the CT scan, not the patient, is the highest auditable clustering unit.