BiCC: Bidirectional Connected-Component Loss for Instance-Aware Segmentation

AI summary

Overview

Research area: Medical image segmentation, specifically instance-aware (lesion-wise) loss functions for 3D segmentation in MRI, CT, and PET/CT.

Technical level: Advanced. The paper assumes familiarity with segmentation loss functions, connected-component analysis, Voronoi partitions, and gradient behavior.

Scope: The paper introduces BiCC (bidirectional connected-component loss), which combines annotation-derived and prediction-derived instance partitions so that both missed lesions and false-positive components receive equally weighted training terms, and evaluates it on five 3D medical datasets using nnU-Net.

What This Paper Is About

Common segmentation losses such as cross-entropy, Dice, and DiceCE aggregate errors voxel-wise, so a lesion's influence on the objective is proportional to its volume and small but clinically important lesions get little weight. Existing instance-aware losses (blob loss, CC-DiceCE) give each lesion its own term, but they derive their regions purely from the annotations, so a false-positive prediction component never receives its own instance-level penalty. BiCC addresses this by pairing a reference-derived partition with a prediction-derived one, so each predicted component is scored on its own spatial scale regardless of size, with a balance parameter α that controls the lesion-wise precision-recall trade-off.

Key Contributions

1. A precision-oriented instance branch built on a prediction-derived partition, including a scoring rule (a mass-weighted mean of predicted probability with stop-gradient weights) for unambiguous false-positive cells whose total gradient magnitude is independent of component size.
2. A controlled five-dataset comparison across MRI, CT, and PET/CT against DiceCE, blob loss, and CC-DiceCE.
3. An analysis of how the balance parameter α allows control along the lesion-wise precision-recall frontier, including comparisons with post-hoc thresholding of a DiceCE model.

Main Findings

• Precision and recall behavior: CC-DiceCE raises recall on every dataset but lowers precision on four, by up to 0.363. BiCC instead increases precision over DiceCE on all five datasets, significantly on three. Recall increases on BraTS-METS and LiTS and decreases on AutoPET III and ISLES, with a non-significant decrease on CMB.
• Lesion-wise F1: BiCC's F1 gains over DiceCE are significant on AutoPET III, BraTS-METS, and LiTS, while ISLES and CMB remain comparable to DiceCE. BiCC outperforms CC-DiceCE in lesion-wise F1 on four datasets and blob loss on all five. Directly compared with CC-DiceCE, BiCC has significantly higher precision and F1 on all four non-CMB datasets.
• False positives on lesion-free scans: On empty-GT scans BiCC produces the fewest predicted components on all five datasets, with significant reductions against both DiceCE and CC-DiceCE on four. For example, on AutoPET III, BiCC reports FP/neg of 3.4293 versus 5.5009 for DiceCE and 10.8482 for CC-DiceCE.
• Global overlap (CC-Dice): Compared with DiceCE, BiCC gains up to +0.014 in CC-Dice (AutoPET III) and gives up at most 0.009 (ISLES), but remains below CC-DiceCE on all five datasets, an ordering the authors attribute to CC-Dice's recall bias.
• Global Dice: Relative to DiceCE, global Dice changes by −0.011 to +0.019, which the authors treat as secondary because global Dice is dominated by large lesions.
• Ablation on α: Across the α sweep from 0 to 1, precision increases monotonically and recall decreases on both ISLES and AutoPET III. At α = 1/4, about 70% of the total precision change and almost 90% of the maximum F1 improvement are already attained, indicating α = 1/2 is a robust default without dataset-specific tuning.
• False-positive scoring ablation: Replacing the mass-weighted score with smoothed Dice (ε = 10⁻⁵) at α = 1/2 produces results closely resembling the α = 0 (CC-DiceCE) operating point on both datasets, confirming that with a small smoothing constant the gradient for unambiguous false positives is effectively zero.
• Post-hoc thresholding comparison: Shifting DiceCE logits for logit shifts s ∈ {−5, …, 5} cannot replicate tuning α. On ISLES, thresholding reaches a comparable peak F1 but spans a substantially narrower precision-recall range, and at high thresholds precision reverses rather than continuing to improve. On AutoPET III, the α sweep strictly dominates thresholding across their shared recall range.
• Gradient size-dependence: In the synthetic experiment of Figure 1, over a 40 × 40 volume sweep, the ℓ₁ norm of instance-loss probability gradients on false-positive cells has its spread reduced from 53.4× (plain Dice) to 5.9×.
• Computational cost: Benchmarked on an NVIDIA RTX A6000, BiCC adds a 17–23% training overhead per epoch over DiceCE and 7–10% over CC-DiceCE, with no additional inference cost.
• Limitation on CMB: On CMB, CC-DiceCE has higher mean F1 than BiCC, though the difference is not significant. CMB has the smallest, most uniform lesions in the study, so CC-DiceCE may already penalize false positives sufficiently through reference-cell overlap.

Methodology in Plain English

Each training patch is partitioned into Voronoi cells twice: once using the 26-connected components of the ground truth, and once using the components obtained by thresholding the predictions at 0.5. In each case, every voxel is assigned to its nearest component under Euclidean distance in physical space. The regional loss is DiceCE restricted to a given Voronoi cell, and each component contributes one term with weight 1/|C| regardless of its volume. If a partition contains no components, the method falls back to whole-patch DiceCE.

For prediction-derived cells that contain no annotation at all (unambiguous false positives), smoothed Dice is nearly flat, so the authors replace Dice in those cells with a mass-weighted mean of the predicted probability plus cross-entropy. The weights are computed from the current predictions but treated as constants during backpropagation (stop-gradient), which makes the derivatives sum to one and removes the dependence of total gradient magnitude on component size. Cells that mix annotation and prediction signals keep DiceCE, applying the precision penalty only to unambiguous false positives.

The full objective adds a bracketed instance term to DiceCE, weighting the reference-derived term by (1 − α) and the prediction-derived term by α, with the bracketed term weighted 1:1 against the global term as in prior component-based objectives. Since that budget is fixed, α controls only the relative contribution of the two partitions.

All methods use the same self-configured 3D nnU-Net and differ only in objective. Each method-dataset combination uses five-fold cross-validation and 500 epochs, with α = 0.5 and ε = 0 to keep methods comparable (the DiceCE baseline did not converge with ε = 10⁻⁵ on CMB). AutoPET III and BraTS-METS folds are grouped by patient to prevent leakage. Metrics are global Dice, CC-Dice, and lesion-wise precision, recall, and F1, with reference and predicted components matched one-to-one at maximum cardinality when IoU > 0.1, all macro-averaged over scans. Empty-GT scans are excluded from those five metrics, and FP/neg is reported separately for them. Comparisons use paired Wilcoxon signed-rank tests with Holm-Bonferroni correction applied per dataset-metric across four contrasts.

Datasets: AutoPET III (1,611 scans; produced from 1,014 FDG and 597 PSMA PET/CT studies), ISLES 2026 (1,453 labeled native-space T1-weighted MRI scans of stroke lesions), BraTS-METS (1,294 scans, tumor core as binary target using T1n, T1c, and T2f sequences), LiTS (131 contrast-enhanced abdominal CT scans, liver tumor class only), and CMB (72 scans, the cerebral microbleed task of VALDO).

Why This Matters

Impact on research: The paper reframes instance-aware segmentation losses as a two-sided problem: misses and false positives each deserve their own equally weighted term. It shows analytically and experimentally why annotation-derived partitions alone leave false positives size-dependent, and it provides a controllable knob (α) for moving along the lesion-wise precision-recall frontier at training time rather than at inference time.

Real-world applications:

• Computer-assisted review of whole-body PET/CT tumor lesions (AutoPET III), where each false-positive component may require separate inspection.
• Brain metastasis detection and delineation on multisequence MRI (BraTS-METS), where missing a lesion is serious but excessive false positives add review burden.
• Liver tumor segmentation in contrast-enhanced abdominal CT (LiTS).
• Cerebral microbleed detection (CMB/VALDO) and stroke lesion segmentation on T1-weighted MRI (ISLES 2026).

Industry relevance: BiCC requires no architectural changes and incurs no additional inference cost, making it a practical drop-in loss for existing nnU-Net-based pipelines. Its ability to reduce false-positive components on lesion-free scans, and to let a user choose the precision-recall trade-off appropriate to a use case, maps directly onto deployment concerns in clinical screening and review workflows. Code is available at https://github.com/TIO-IKIM/BiCC-Loss.

Future Directions

• Handling "mixed cells," where a prediction-derived cell contains annotation that does not overlap its generating component; separating those signals would require rules for component support, matching, splits, and merges, which the current method avoids by retaining DiceCE in those cells.
• Determining how to select α automatically or per dataset, since the authors use a fixed α = 0.5 without per-dataset tuning and only identify α = 1/2 as an empirically robust default.
• Understanding the CMB case, the one dataset where CC-DiceCE has higher mean F1 than BiCC and where none of the precision, recall, or F1 differences is significant, to establish which lesion profiles the prediction-derived branch may over-penalize.
• Testing whether the bidirectional connected-component formulation transfers beyond 3D medical segmentation, given that it requires no architectural changes.

Target Audience

Researchers and practitioners working on 3D medical image segmentation and instance-aware loss design, particularly those using nnU-Net or similar pipelines; method developers interested in gradient behavior of region-based losses; and clinical-AI engineers who need to trade lesion-wise precision against recall and reduce false-positive review burden. Some familiarity with Dice and cross-entropy losses, connected components, and Voronoi partitions is assumed.

Authors’ abstract

Common segmentation losses aggregate errors voxel-wise, so lesions influence the objective in proportion to their volume, giving small but clinically critical lesions disproportionately little weight. Instance-aware losses aim to address this mismatch by assigning each lesion its own term. However, blob loss and CC-DiceCE derive their regions solely from annotations, so false-positive components receive no instance-level term. This matters in computer-assisted review, where each false-positive component may require separate inspection, making precision and false-positive burden important alongside recall. We introduce the bidirectional connected-component loss (BiCC), which pairs annotation- and prediction-derived partitions to score predicted components on their own scale. By deriving instances from the predictions, this branch directly penalizes false-positive components regardless of their size. The balance parameter $α$ allows control over the lesion-wise precision-recall trade-off. Across five datasets with five-fold cross-validation using nnU-Net, BiCC outperforms CC-DiceCE in lesion-wise F1 on four datasets and blob loss on all five. It significantly improves over DiceCE on three datasets and matches it on two; CC-DiceCE instead loses up to 0.363 precision by favoring recall. Code is available at https://github.com/TIO-IKIM/BiCC-Loss.