Iterative pseudo-labeling based adaptive copy-paste supervision for semi-supervised tumor segmentation

• Expert tumor annotation is labor-intensive and variable, motivating SSL, which uses both labeled and unlabeled data.Source: §1 · p. 2
• Medical-image SSL includes consistency regularization, self-training and pseudo-labeling, and adversarial methods.Source: §1 · p. 2
• IPA-CP uses student–teacher prediction divergence to guide adaptive augmentation and iteratively aligns student pseudo labels with teacher predictions.Source: §1 · pp. 2–3
• Prior SSL tumor work largely targets large organs, leaving small-tumor segmentation underaddressed despite its clinical importance.Source: §1 · pp. 2–3
• IPA-CP integrates labeled and unlabeled data without extra network components or training losses, using uncertainty-adaptive augmentation and iterative pseudo-label transition.Source: §1 · p. 3
• Experiments use a newly constructed in-house FSD dataset and public MSD data to assess small-tumor segmentation and generalization across tumor types.Source: §1 · p. 3

2.1. Fully supervised medical tumor segmentation

• Fully supervised tumor segmentation includes U-Net-based and non-U-Net models, but often requires extensive labels and can carry large parameter counts in 3D.Source: §2.1 · p. 3

2.2. Semi-supervised medical image segmentation

• Medical SSL work spans consistency regularization, self-training and pseudo-labeling, and adversarial learning approaches.Source: §2.2 · pp. 3–4
• Many SSL methods target larger organs, while some augmentation-heavy approaches incur high computational costs and lack dynamic adaptation for tumors.Source: §2.2 · p. 4

3. Methodology

• IPA-CP combines two-way uncertainty-based adaptive augmentation, iterative pseudo-label transition, and bidirectional copy-paste supervision for learning from labeled and unlabeled data.The framework is described as using these three modules without additional parameters or modules.Source: §3 · p. 4

3.1.1. Prediction disagreement generation

• Weak and strong views of unlabeled samples, including randomly masked strong views, generate student and teacher predictions for measuring disagreement.The mean-teacher architecture is used to model prediction discrepancies.Source: §3.1.1 · pp. 4–5

3.1.2. Two-way uncertainty estimation

• Two-way KL divergences measure student-to-teacher and teacher-to-student prediction differences, with greater divergence indicating greater uncertainty.The method estimates inconsistency between the two models because unlabeled data lack precise ground-truth labels.Source: §3.1.2 · pp. 5–6

3.1.3. Adaptive augmentation

• Adaptive augmentation uses the high-uncertainty score to make low-uncertainty samples resemble weak views and other samples resemble masked strong views.The uncertainty score filters voxels using τ = 0.9.Source: §3.1.3 · p. 6

3.2. Iterative pseudo-label transition

• Iterative pseudo-label transition moves perturbed student predictions toward teacher predictions, producing more robust and diverse pseudo-labels for unlabeled samples.The authors motivate refinement because the teacher does not always outperform the student on unlabeled data.Source: §3.2 · p. 7

3.3. Bidirectional copy-paste supervision

• Bidirectional copy-paste swaps background regions between labeled and unlabeled samples for paired supervision, while omitting largest-component filtering to retain disconnected tumor fragments.The method uses refined pseudo-labels for unlabeled samples and processes labeled ground truth similarly.Source: §3.3 · p. 8

3.4. Loss function

• The student network is trained with a combination of cross-entropy and Dice losses.Source: §3.4 · p. 8

4.0.1. FSD

• The in-house FSD dataset contains 558 patients; 390 scans were allocated to training, 168 to testing, and 30 training samples to validation.Esophageal tumors were manually delineated and verified by two radiotherapists.Source: §4.0.1 · p. 8

4.0.2. MSD

• Of 200 LiTS scans, 116 tumor-containing scans were retained and split into 100 training, 8 validation, and 8 testing scans.The original LiTS test ground truth is unreleased; the study used its own split of the retained scans.Source: §4.0.2 · pp. 8–9
• FSD and MSD intensities were truncated to dataset-specific HU ranges, resampled to standard voxel resolutions, and min-max normalized.FSD was resampled to 0.8 × 0.8 × 5 mm3 and MSD to 2 × 2 × 2 mm3.Source: §4.0.2 · p. 9

5.1. Implementation details and evaluation measures

• IPA-CP was implemented in PyTorch and MONAI with a 3D U-Net base network and an NVIDIA RTX 3090 GPU.The batch size was 2, with patch sizes of 160×160×80 for esophageal tumors and 112×112×64 for liver tumors.Source: §5.1 · p. 9
• Training used Adam with a polynomial learning-rate policy, 6,000 iterations for esophageal tumors, 42,000 for liver tumors, and EMA decay α = 0.99.Reported performances are averages of 3 runs.Source: §5.1 · p. 9
• Evaluation metrics were Dice, Jaccard index, 95% Hausdorff distance, root mean squared error, and average surface distance.Source: §5.1 · p. 9

5.2.1. FSD

• 7.95% higher Dice than the baseline is reported for IPA-CP with 10% labeled data on FSD.The authors also report performance close to fully supervised learning with 100% labeled data.Source: §5.2.1 · p. 9, Table 2 · p. 10

5.2.2. MSD

• On MSD, IPA-CP is competitive in Dice, JA, and RMSE, while SDCL and DACNet perform well on FSD but fail on tumor segmentation.The authors report that IPA-CP’s results on MSD support generalization to datasets with small tumors.Source: §5.2.2 · pp. 9, 11

5.2. Quantitative evaluation and comparison

• IPA-CP achieves the highest Dice and JA scores across both FSD and MSD, and leads on all five evaluation metrics at both labeled-data ratios.The comparison uses 10% and 20% labeled data.Source: §5.2.1 · p. 9, Table 2 · p. 10, §5.2.2 · p. 11

5.2.2. MSD

• Increasing the labeled-data proportion improves performance overall across the compared methods.The authors report this pattern in their observations of Table 2.Source: §5.2.2 · p. 11

5.3. Qualitative comparison

• For the illustrated esophageal tumor case, IPA-CP aligns closely with ground truth, while other methods produce false positives or miss tumor regions.This is the relatively large-tumor example in the second row.Source: §5.3 · p. 11
• On the illustrated MSD cases, IPA-CP identifies small tumors that other SSL methods fail to detect or segment accurately.The fourth and fifth rows show small tumors and incorrect voxel predictions by other methods.Source: §5.3 · p. 11
• Across the visualized cases, other methods more often segment non-target regions or fail to identify all tumors.The authors attribute the difficulty to ambiguous CT tumor voxels that vary in size.Source: §5.3 · p. 11

5.4.1. FSD

• 71.72% Dice and 7.29 mm HD are achieved on FSD when IPT and PDI are added, the best overall results in the reported ablations.The authors describe TUE and PDI as critical for refining predictions on FSD.Source: §5.4.1 · p. 12
• On FSD, adding BCP raises Dice from 65.51% to 70.61% and reduces HD from 11.09 mm to 7.63 mm.The authors associate this gain with increased training-sample diversity for large tumors.Source: §5.4.1 · p. 12

5.4.2. MSD

• 46.61% Dice and 37.34% JA are achieved by the full model on MSD, the best result reported there.The authors identify prediction disagreement filtering as essential for small, uncertain tumor regions.Source: §5.4.2 · p. 12
• On MSD, TUE raises Dice to 43.98% and reduces HD to 21.37 mm, while IPT without TUE yields a Dice score of 38.22%.The authors report that TUE guides ambiguous regions and IPT provides additional guidance for small tumors.Source: §5.4.2 · p. 12
• The authors attribute the stronger improvements on MSD than FSD to MSD’s smaller tumors and note that BCP-based methods benefit from FSD’s larger tumors.They relate the latter benefit to BCP’s reliance on the largest connected component strategy.Source: §5.4.2 · p. 12

5.5.1. Effects of hole number and size in M

• More than 30 holes reduced MSD performance, whereas FSD remained stable; hole counts were randomized from 5/10 to 40/30 for FSD/MSD.Hole sizes were set to 10/10 through 40/20 for FSD/MSD.Source: §5.5.1 · p. 13

5.5.2. Effects of threshold

• A threshold τ of 0.9 produced optimal performance on both datasets, filtering uncertain voxels and retaining more informative ones for augmentation.Source: §5.5.2 · p. 13

5.5.3. Effects of pseudo labeling strategies

• IPT alone substantially improved MSD and slightly improved FSD, while combining VOT with IPT further enhanced FSD performance.VOT is used for the first 1/5 of training iterations, followed by IPT for the remaining 4/5.Source: Table 5 · p. 13, §5.5.3 · p. 14

5.5.4. Effects of loss function

• 46.61% Dice on MSD with CE+Dice (vs. 43.00% with CE) and 71.72% on FSD (vs. 71.14% with CE) were the best tested results.Dice loss scored 12.76% on MSD and 68.63% on FSD.Source: §5.5.4 · p. 14, Figure 5 · p. 14

5.5.5. Limitation and future work

• IPA-CP can still make false predictions on extremely small or barely visible tumors, and its performance beyond CT has not been established.Future work proposes context-aware features with spatial and structural priors for tiny tumors, and modality-specific adaptations and augmentation for MRI and ultrasound.Source: Figure 6 · p. 14, §5.5.5 · p. 15

• IPA-CP uses two-way uncertainty-based adaptive augmentation to increase the diversity and informativeness of unlabeled data.Source: §6 · p. 15
• An iterative pseudo-label transition strategy guides predictions toward more robust pseudo labels for unlabeled samples.Source: §6 · p. 15
• Bidirectional copy-paste facilitates learning from both labeled and unlabeled data, and the approach can be integrated into other models.Experiments on the in-house esophageal and public liver tumor datasets demonstrate the method's efficacy.Source: §6 · p. 15