Source-linked AI summary
ConPro: Contrast Projection Pretraining for Label-Efficient Vessel Segmentation in DSA Sequences
Xinge Guo, Yuanhao Wang, Liqi Shu, Yang Liu, Min Xu
TL;DR
Dense vessel masks require labor-intensive pixel-by-pixel annotation, while semi-supervised methods derive targets from the current model and common self-supervised pretexts reconstruct appearance rather than using DSA contrast passage. ConPro pretrains a backbone to predict C⋆, the normalized per-pixel drop below the temporal median, from unlabeled DSA sequences before transferring it to segmentation. Compared with supervised training, ConPro has a higher three-seed mean Dice and clDice at every label fraction on DIAS and DSCA, with gains of 4.4 Dice and 5.2 clDice at DIAS 10%.
Problem
Dense vessel masks require labor-intensive pixel-by-pixel annotation, while semi-supervised methods derive targets from the current model and common self-supervised pretexts reconstruct appearance rather than using DSA contrast passage.
Source: §1 · p. 1Method
ConPro pretrains a backbone to predict C⋆, the normalized per-pixel drop below the temporal median, from unlabeled DSA sequences before transferring it to segmentation.
Source: §1 · p. 1, §2.1 · p. 2Results
Compared with supervised training, ConPro has a higher three-seed mean Dice and clDice at every label fraction on DIAS and DSCA, with gains of 4.4 Dice and 5.2 clDice at DIAS 10%.
Source: §4.1 · p. 4Takeaways & Limitations
UniMatch initialized with ConPro weights gains 0.5 to 2.0 Dice and 0.9 to 2.3 clDice at every label fraction without changing the segmentation architecture.
Source: §4.1 · p. 4Motion artifacts enter C⋆ and vessels that never opacify are absent; behavior at true lesions and clinical value remain unmeasured.
Source: §4.3 · p. 4
Abstract
from arXiv · showhide
Dense vessel annotation in digital subtraction angiography (DSA) is labor-intensive, yet every unlabeled sequence records how contrast passes through the vessels. Semi-supervised methods take their targets from the current model, and generic self-supervised pretexts reconstruct static appearance, so this signal goes unused. We propose ConPro, a self-supervised pretraining scheme whose target is a contrast projection, the normalized drop of every pixel below its temporal median over the sequence. On DIAS and DSCA, with 10%, 20% and 50% of the training cases labeled, ConPro improves on training from scratch at every label fraction and is the best of the compared methods on DSCA at 20% and 50% labels. Controlled comparisons show that the gain comes from the target. A temporal-median target with the same input, loss and budget stays at scratch level, and using the projection directly instead of learning it, as an input channel or a pseudo-label, helps little or hurts. ConPro provides pretrained weights without changing the segmentation architecture, so it combines with semi-supervised training, and UniMatch, the strongest baseline, gains 0.5 to 2.0 Dice and 0.9 to 2.3 clDice at every label fraction when started from ConPro weights, reaching 75.4 Dice on DIAS and 81.3 on DSCA.
1. INTRODUCTION
DSA sequences record contrast passage, but dense vessel masks require pixel-by-pixel annotation and expert review. Existing semi-supervised methods derive targets from the current model, while self-supervised methods reconstruct appearance.
1. INTRODUCTION
- Only 60 and 224 labeled sequences are available in the two public datasets, despite DIAS being selected from over a thousand acquisitions.Dense masks are drawn pixel by pixel and reviewed by senior neurosurgeons.Source: §1 · p. 1
- Semi-supervised methods derive pseudo-labels or consistency targets from the current model, while self-supervised pretraining uses targets such as masked appearance or missing pixels.The passage notes that targets from the current model are bounded by that model.Source: §1 · p. 1
1 T/2 T Frame index t
ConPro uses contrast passage in DSA sequences as a mask-free pretraining target, then transfers the pretrained backbone to vessel segmentation. Controlled comparisons attribute the gain to learning to predict that target.
1 T/2 T Frame index t
- Angiographic pretexts had reconstructed appearance, learned subtraction, or extracted vessels, while leaving the temporal course of contrast aside as a target..Source: §1 · p. 1
- ConPro defines C⋆ as each pixel’s largest drop below its temporal median and uses it as a soft-mask target for pretraining before artery segmentation with few labels.The target turns sequence opacification into supervision without manual annotation.Source: §1 · p. 1
- Controlled comparisons hold input, loss, and budget fixed while swapping targets, vary frame selection, and test C⋆ as an input channel or pseudo-label; together they place the gain on learning to predict the target.The pretrained weights also serve as a drop-in initialization for semi-supervised methods, tested with UniMatch and RPST∗.Source: §1 · p. 2
- Pretraining uses unlabeled angiography sequences, and inference keeps the segmentation network unchanged, adding no parameters, inputs, or computation.The largest reported gain is 4.4 Dice and 5.2 clDice with only three labeled DIAS sequences.Source: §1 · p. 2
2. METHOD
ConPro first predicts a contrast projection from unlabeled DSA frames, then transfers the backbone to supervised vessel segmentation. Its target encodes transient darkening relative to each pixel’s temporal median.
2. METHOD
- ConPro has two stages: predict C⋆ from unlabeled sequences, then transfer the backbone to supervised vessel segmentation.The Stage II network uses the pretrained backbone with a segmentation head.Source: §2 · p. 2, §2.3 · p. 2
2.1. Contrast projection target
- The contrast projection encodes each pixel’s strongest darkening relative to its temporal median, then normalizes sequence values using the 99th percentile of R over valid frames before padding.The median reference tolerates a first frame that already carries contrast.Source: §2.1 · p. 2
2.2. Stage I: pretraining objective
- Stage I predicts the target from input frames that may be all frames or a uniformly sampled, temporally ordered subset; the default uses a random half as input and the full-sequence target.The BCE-plus-Dice loss accepts continuous targets.Source: §2.2 · p. 2
2.3. Temporally pooled backbone and Stage II transfer
- TPoolUNet uses a shared 2D encoder, masked temporal max pooling, and a U-Net decoder; Stage II initializes the backbone from Stage I and fine-tunes it with labeled masks.Stage II adds a randomly initialized 1 × 1 segmentation head and uses BCE plus Dice.Source: §2.3 · p. 2
3. EXPERIMENTAL SETUP
Experiments use nested labeled subsets from DIAS and DSCA, with a shared backbone and matched Stage II update budget across segmentation methods. Stage I ablations separately vary the pretraining target or frame input, and performance is measured by sequence Dice and clDice.
Data and evaluation
- Evaluation uses nested 10%, 20%, and 50% labeled subsets: 3/6/15 of 30 DIAS training cases and 15/31/76 of 153 DSCA non-validation cases.Stage I uses 90 DIAS training and unlabeled sequences or 153 DSCA training sequences; DIAS sequences retain 4–14 frames and DSCA sequences are resampled to eight.Source: §3 · pp. 2–3
Compared methods
- Table 1 reports Dice/clDice three-seed means for supervised, semi-supervised, and self-supervised methods, including direct-use C⋆ controls and methods initialized from ConPro weights.The table uses a matched backbone and Stage II update budget.Source: Table 1 · p. 3
Ablations
- Stage I ablations use independent protocols: Target fixes input to a random half and varies the target, while Input keeps the contrast projection and varies frame feeding.Stage I runs 8,000 AdamW updates; Stage II runs 1,200 AdamW updates, with validation every 150 updates.Source: Table 2 · p. 3, §3 · p. 4
Data and evaluation
- Results are sequence Dice and topology-aware clDice, macro-averaged over test cases.Validation selects the Stage II checkpoint, which is evaluated once on the test set.Source: §3 · p. 3
Compared methods
- All segmentation methods use TPoolUNet and share labeled subsets and Stage II update budgets; CPS, CorrMatch, UniMatch, RPST*, and Generic-SSL provide comparison methods or controls.ConPro weights also initialize UniMatch and RPST*; Generic-SSL predicts the all-frame temporal median from a patch-masked sequence with an L1 loss.Source: §3 · pp. 3–4
4. RESULTS
ConPro improves on supervised training across label fractions and datasets, with the strongest DSCA comparisons favoring ConPro or UniMatch initialized from its weights. Ablations favor learning the contrast-projection target, while the reported scope excludes lesion behavior and clinical value.
4.1. Performance with limited labels
- ConPro exceeds supervised training on both metrics at every label fraction on DIAS and DSCA, and leads UniMatch on DSCA at 20% and 50% labels by 1.4–2.0 Dice and 1.3–2.1 clDice.At DIAS 10%, the largest gains are 4.4 Dice and 5.2 clDice; the DSCA leads over UniMatch are significant at p < 0.01.Source: §4.1 · p. 4
- UniMatch initialized with ConPro weights gains 0.5–2.0 Dice and 0.9–2.3 clDice over UniMatch from scratch at every label fraction.The combination is the table’s best entry on DIAS and at DSCA 10%.Source: §4.1 · p. 4
4.2. Effect of target and input
- Using the same random half as input and BCE+Dice loss, all-frame-median prediction leaves every mean within 0.1 of scratch or below; C⋆ raises Dice by 1.2–5.2 and clDice by 1.4–5.9 over that target.Source: §4.2 · p. 4
- Input variants differ little overall; averaged over label fractions, the default beats all-frame input by 1.0 Dice and 1.1 clDice on DIAS, while DSCA variants remain within 0.3 points.The default uses half the frames with the full target; all-frame training is 0.5 points below scratch at DIAS 20%.Source: §4.2 · p. 4
- Using C⋆ directly as an input channel or pseudo-label helps little or hurts, favoring its use as a pretraining target in the tested controls.Thresholded C⋆ reaches 52.0 and 59.3 Dice; pseudo-labeling raises DIAS 10% clDice from 58.7 to 63.0 but costs two to four points on DSCA.Source: §4.2 · p. 4
4.3. Scope and limitations
- Because frame pooling makes shuffling an identity, the evidence concerns which contrast phases are seen, not their order; motion artifacts enter C⋆, and never-opacifying vessels are absent.Behavior at true lesions and clinical value remain unmeasured.Source: §4.3 · p. 4
5. CONCLUSION
ConPro uses unlabeled DSA sequences to pretrain a backbone on vessel-localized contrast change, then fine-tunes it for segmentation. The conclusion reports gains over scratch across label fractions and describes the weights as compatible with semi-supervised training.
6. CONCLUSION
- ConPro trains a backbone to predict the per-pixel drop below the temporal median from unlabeled DSA sequences, then fine-tunes it with available labels.The conclusion describes the drop as marking where contrast arrived.Source: §5 · p. 4
- ConPro improves over training from scratch at every label fraction, is best among compared methods on DSCA at 20% and 50%, and improves UniMatch when initialized with its weights.Source: §5 · p. 4
- The conclusion attributes the ablation gains to the target itself and notes that the method uses data already present in angiography archives.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The retrospective study used open-access human-subject data from DIAS and DSCA; dataset licenses confirmed that ethical approval was not required.