SynSeq: End-to-End SYNTAX Score Prediction from Coronary Angiography Videos

1 Introduction

• Manual SYNTAX scoring is time-consuming and subject to inter-observer variability, limiting its feasibility in everyday practice.The score is used to grade lesion severity, complexity, and location and inform revascularization decisions.Source: §1 · pp. 1–2
• Prior automated angiography methods largely address component tasks or binary disease discrimination, leaving continuous SYNTAX prediction across complex cases comparatively unexplored.Most earlier methods use single frames, and the CardioSyntax baselines emphasize zero-versus-nonzero classification.Source: §1 · p. 2
• SynSeq directly optimizes continuous patient-level SYNTAX scores from multi-view coronary angiography videos using targeted preprocessing and a tailored strategy for skewed labels.The strategy combines sample weighting with linear rather than logarithmic target scaling to preserve granularity in high-severity cases.Source: §1 · pp. 2–3
• Separate left- and right-vessel branches each encode angiography sequences and predict a SYNTAX score and vessel status; inference sums their score predictions.The branches share an architecture but have separate weights.Source: Figure 1 · p. 3
• Evaluation on the full CardioSyntax dataset reports significantly improved performance over prior state-of-the-art methods and increased robustness across multiple expert annotations.The evaluation uses the full dataset and compares performance across expert annotations.Source: §1 · p. 3

• SynSeq processes left and right coronary trees independently, averages per-video sub-score predictions, then sums nonnegative side scores into a patient-level prediction.Source: §2 · pp. 3–4
• Angiography-specific training augmentations mimic acquisition variation; each patient must have at least one labeled video per tree, and clips use temporal center cropping or repetition padding.Source: §2 · p. 4
• Training addresses zero-inflated, skewed scores with inverse-frequency sample weighting and linear [0, 1] target scaling in weighted MSE rather than logarithmic scaling.Source: §2 · p. 4
• A multi-task objective combines regression with sample-weighted binary cross-entropy for disease-presence classification, balanced by hyperparameter α.Source: §2 · pp. 4–5
• Each tree's backbone is pretrained per video, then frozen while the dual-head LSTM is trained, before joint fine-tuning; sub-scores are aggregated at inference.Source: §2 · p. 5

• 1,844 patients in CardioSyntax contribute 9,590 left-artery and 3,970 right-artery videos, with primary sub-scores and additional consensus labels for 60 patients.Source: §3 · p. 5
• SynSeq is compared with MVL, a single-view DeepRV linear-probing baseline, and a naive mean using a patient-wise split with 60 multi-expert patients held out for testing.The remaining data use five-fold cross-validation, and test performance is reported as mean ± standard deviation across folds.Source: §3 · pp. 5–6
• Regression is assessed with R2, MAE, and mean-error bias, while treatment recommendations use weighted F1 across low, intermediate, and high risk categories.Source: §3 · p. 6
• Performance sensitivity is tested with a second 80/20 SYNTAX-stratified split of the full cohort, while the 60-patient consensus test split is the default.Source: §3 · p. 6
• Ablations cumulatively remove sample weighting, linear target-space optimization, or domain-specific preprocessing to measure changes in R2, MAE, and bias.Source: §3 · p. 6
• The implementation uses a Kinetics-400-pretrained R3D-18 backbone and projected LSTM, with backbone pretraining, frozen-backbone LSTM training, and joint fine-tuning.Source: §3 · p. 6

4 Results

• R² was 0.62 for SynSeq versus 0.07 for MVL and −0.11 for DeepRV on Expert 1 test labels.Table 1 reports regression performance on the test set using Expert 1 labels.Source: Table 1 · p. 7, §4 · p. 7
• For diseased samples, SynSeq reached R² 0.54, bias −0.90, and MAE 8.24, versus MVL/DeepRV R² −0.14/−0.17, bias −10.47/−2.28, and MAE 11.90/14.09.Improvements were consistent across folds and significant for bias (p<0.001) and MAE (p<0.05).Source: §4 · p. 7
• Weighted F1 was 0.80 ± 0.04 for SynSeq versus 0.64 ± 0.05 for MVL; SynSeq misclassified 11/60 patients versus 17/60.Only 3 SynSeq errors crossed multiple decision thresholds in a way that would alter treatment strategy, versus 12 for MVL.Source: §4 · p. 7, Figure 2 · p. 8
• Across three experts, SynSeq’s R² stayed 0.60–0.62, near expert-pair agreement of 0.65–0.67 but below the 0.83 maximum.Figure 2(b) shows reduced error variability versus MVL, approaching variability between human experts.Source: §4 · pp. 7–8, Table 2 · p. 7
• Removing sample weighting, linear scaling, or preprocessing degraded performance; without zero-clamping, R² was 0.58 ± 0.07, bias −2.15 ± 1.23, and MAE 7.87 ± 1.17.Zero-clamping affected an average of 16.6 of 60 test samples per side across folds.Source: §4 · p. 8, Table 3 · p. 8
• After full training, the healthy-versus-diseased head reached macro F1 0.747 ± 0.047 and accuracy 0.783 ± 0.051, above MVL at 0.703 ± 0.020 and 0.743 ± 0.023.Source: §4 · p. 9

• SynSeq consistently reduces error and improves agreement over MVL, with stable performance across three annotators despite training on one expert's labels.The authors suggest this reflects alignment with angiographic patterns rather than overfitting to individual annotation styles.Source: §5 · p. 9
• CardioSyntax has no official split, so its reported results may estimate different quantities and are not directly comparable with SynSeq's evaluation.The original study used five-fold cross-validation without a separate held-out test set; the authors used a held-out cohort and multiple seeds.Source: §5 · p. 9
• The baseline comparison is limited because CathAI, DeepCoro, and Deep-CoroCLIP weights and training data are unavailable.The authors also report that natural-prevalence class imbalance affects model calibration and performance.Source: §5 · p. 9
• External validation across institutions and imaging protocols is still required to assess generalizability beyond CardioSyntax.Source: §5 · p. 10
• The authors identify domain-specific preprocessing, linear-space optimization, and explicit sample weighting as design choices supporting improved and stable prediction.They connect linear-space training to the additive structure of SYNTAX scores and sample weighting to reduced underestimation near clinical decision thresholds.Source: §5 · pp. 9–10
• SynSeq has the potential to support more consistent and scalable cardiovascular treatment planning.Source: §5 · p. 10