acceptodds
Withdrawn from ICLR 2027

Two-Sided Safe Screening for Scalable Tversky Optimization in Segmentation Decoding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image segmentation is a fundamental task in computer vision. Segmentation networks typically output pixel- or voxel-level probabilities, whereas overlap measures evaluate discrete prediction masks. Metric-aware decoding uses these probabilities to optimize overlap measures, but exact optimization of the resulting decoding objective can be computationally demanding for large images and volumes. We study scalable decoding for a broad family of Tversky-type objectives, including Dice, intersection over union, and -scores, under a plug-in expected linear-fractional (ELF) criterion defined by a ratio of expected confusion counts. We characterize the ranking-induced structure of its maximizers and derive a pair of complementary, closed-form safe-screening certificates: an upper-threshold certificate identifies coordinates that must be selected, whereas a data-adaptive lower-threshold certificate identifies coordinates that must be excluded before downstream exact optimization. Together, these two-sided certificates reduce the effective problem dimension while preserving a global optimizer of this criterion. We instantiate the screening rules in an efficient GPU implementation with a sorting based downstream backend. Experiments on synthetic probability maps and native medical volumes demonstrate substantial reductions in runtime and working memory at large scales.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Safe screening, Image segmentation, Tversky index, Metric-aware decoding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Zf9EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work adapts the safe screening principle to metric decoding for Tversky-based segmentation. Prior to sorting and prefix search, closed-form certificates partition candidate coordinates into three disjoint categories: must-select, must-reject, and uncertain. Precise optimization is carried out solely on the uncertain subset, with the global optimality of the resulting solution fully preserved.

Critical strengths

The paper presents solid novelty and practical value. In particular, sorting and memory bottlenecks are real challenges for 3D medical volumes. The result showing that 99.99% of coordinates can be screened out is compelling.

Critical weaknesses

The adoption of three reference prefixes may be overly conservative.

Questions for authors

The bound derived in Corollary 1 appears overly conservative. Additionally, the statement of Corollary 2 is not sufficiently clear and requires further elaboration.

Reviewer_AgzoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes safe screening rules to accelerate Tversky-based segmentation decoding. The rules identify pixels or voxels that can be fixed as included or excluded, leaving only the unresolved set for exact optimization. They preserve a global optimum of the stated ELF objective, which is a ratio of expected confusion counts rather than the expected segmentation metric itself. Experiments on synthetic probability maps and full medical volumes demonstrate reductions in decoding time and GPU memory use.

Critical strengths
  1. The screening rules reduce computation while preserving a global optimum of the target objective. They also work independently of the downstream solver.
  2. The evaluation on full medical volumes provides convincing evidence of practical memory savings.
Critical weaknesses

The theoretical contribution is relatively incremental, and the added benefit of the two-sided design may be not isolated experimentally.

Questions for authors

How do inclusion-only, exclusion-only, and two-sided screening compare in runtime, memory, and remaining problem size? On the medical volumes, does the inclusion set substantially tighten the exclusion threshold?

Reviewer_GRumOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies computationally efficient decoding for Tversky-type segmentation objectives under a plug-in expected linear-fractional (ELF) criterion, defined as a ratio of expected confusion counts. Its main contribution is a pair of pre-optimization screening certificates: an inclusion threshold fixes sufficiently high-probability coordinates, and an exclusion threshold removes sufficiently low-probability coordinates. Exact optimization is then restricted to the unresolved coordinates while preserving a global optimizer of the ELF objective. The paper evaluates sorting and Dinkelbach backends, separates PyTorch screening gains from additional Triton implementation gains, and reports synthetic and full-volume medical segmentation benchmarks. The main algorithm’s exactness concerns the ELF objective, rather than the exact expected samplewise Tversky score.

Critical strengths

The experiments separate the computational gains from screening from the additional benefits of Triton implementation optimizations. The paper also reports screening overhead on small inputs, providing useful context for interpreting the performance results.

Critical weaknesses

1. Limited methodological novelty

The screening construction builds on an established optimal-threshold formula for linear-fractional metrics. The inclusion certificate follows from an upper bound of one on the optimal score, while the exclusion certificate uses lower bounds obtained from three inexpensive feasible masks. I therefore find the conceptual increment relatively limited: the main contribution is a particular application of existing threshold structure and objective-value bounds to pre-sort screening. Although computationally useful, the manuscript does not sufficiently establish the nontrivial advance over existing results or explain whether the proposed certificates offer substantially stronger guarantees or tighter screening.

2. Incomplete support for mutually exclusive multiclass segmentation

The framework is formulated for binary decoding, and the medical experiments handle multiple foreground classes through independent binary optimizations rather than joint decoding under a mutual-exclusivity constraint. Consequently, the method does not itself guarantee a valid mutually exclusive multiclass label map. If independently decoded masks overlap, additional conflict resolution is necessary and may alter the per-class objective values. The paper should report the frequency of these conflicts, explain how final multiclass predictions are constructed, and evaluate the resulting impact on segmentation quality and computational cost.

3. Unestablished end-to-end deployment benefit

The reported speedups concern decoding and mask reconstruction, excluding network inference, I/O, and host-to-device transfer. For example, lung tumour decoding decreases from 20.86 ms to 2.92 ms, saving approximately 17.94 ms per call; without the decoder’s share of total pipeline latency, the overall application benefit remains unclear. Similarly, memory measurements report additional allocations beyond the resident probability tensor and CUDA context rather than full-pipeline peak memory. A complete pipeline profile, together with evidence of increased throughput, higher concurrency, or larger processable volumes on the same hardware, is needed to substantiate the practical deployment significance.

4. Limited comparison against equivalently optimized alternative solvers

The paper includes relevant Dinkelbach baselines and separates screening gains from additional Triton implementation gains. However, the solver comparison uses standard PyTorch operations and focuses on a highly screenable synthetic distribution. It therefore remains unclear whether the advantage persists against an equivalently optimized sort-free backend across less favorable workloads. Comparing fused, memory-efficient Dinkelbach implementations under identical timing boundaries and output requirements would strengthen the evidence that the proposed certificates provide a robust practical advantage beyond the evaluated implementations.

Questions for authors

See Critical Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.