Coarse-to-fine processing is a widely used strategy for three-dimensional lesion delineation because processing entire high-resolution volumes is computationally expensive while pathology typically occupies only a small fraction of the scan. However, cascaded approaches have a structural failure mode: if a lesion is missed at the coarse stage, it is often absent from the refined region-of-interest and cannot be recovered later. This vulnerability is particularly pronounced for small, low-contrast, and multifocal lesions.
The preprint introduces RECALLREFINE, a coarse-to-fine framework that explicitly treats candidate-region selection as a recall-critical learning problem rather than relying on deterministic cropping around the coarse-stage mask. The central objective is to guarantee that each annotated lesion instance is covered by at least one selected refinement block under a fixed computational budget, thereby reducing irreversible coarse-stage misses.
RECALLREFINE operates in stages. First, a full-volume network processes the entire 3D scan at a coarse resolution and outputs per-voxel coarse lesion probabilities together with multi-scale feature representations. These outputs supply both the rough localization signal and features for subsequent scoring.
A candidate score network examines overlapping 3D blocks extracted across the volume and computes a score for each block. The candidate score combines multiple evidence sources: the coarse probability within the block, a measure of predictive uncertainty, indicators of feature novelty relative to previously-seen regions, and spatial context. This multi-factor scoring is intended to prioritize blocks that are likely to contain lesions that would otherwise be missed by a purely probability-thresholded crop.
Key to the approach is a differentiable selector that chooses a limited set of overlapping blocks for high-resolution refinement under a fixed candidate budget. During training, the selector is optimized with a component-wise coverage objective: the loss requires that every annotated lesion instance remains covered by at least one selected block. This constraint explicitly enforces instance-level recall within the selection process, converting the candidate selection problem into a learnable, recall-focused task.
The selector thus balances competing demands: respecting the fixed refinement budget while maximizing the probability that difficult lesion instances remain represented for native-resolution refinement.
Blocks chosen by the selector are processed by a local refiner at native resolution. The local refiner includes a context-ring branch designed to provide surrounding anatomical context beyond the central block. This helps the refiner produce more accurate local delineations that account for nearby structures.
Because selected blocks overlap, their refined predictions are merged back into the full volume to form the final segmentation. The merging step reconciles overlapping predictions to provide a coherent full-volume output.
The authors evaluated RECALLREFINE on four public 3D lesion benchmarks (the source lists the benchmarks collectively but does not name them individually). Reported quantitative improvements over the strongest matched coarse-to-fine baseline include an increase in mean tumor Dice from 76.5% to 78.4%. For small lesions (below 10 mm), sensitivity improved from 60.9% to 70.8%, indicating a substantial reduction in missed small lesions.
Under a 20% candidate budget—that is, refining only 20% of candidate blocks—the learned selector covered 94.4% of lesion instances, demonstrating the effectiveness of the coverage-constrained selection in preserving instance-level recall despite limited refinement resources.
The results suggest that the primary challenge in cascaded lesion segmentation is not only designing accurate local refiners but ensuring that difficult lesion instances survive the coarse-stage selection. RECALLREFINE reframes candidate selection as a recall-critical learning problem, providing a practical mechanism to trade computational budget for guaranteed instance coverage.
Limitations and specifics that were not reported in the source text: the preprint abstract and metadata summarize methods and aggregated outcomes but do not provide detailed descriptions of network architectures, explicit training hyperparameters, names or characteristics of the four public benchmarks, or runtime and hardware requirements. Those details would need to be obtained from the full paper or supplementary material.
This work is a preprint posted on bioRxiv (doi: https://doi.org/10.64898/2026.09.21.753396) by Wanzhou Chen and Jiayi Chen of Zhengzhou University. The article has not been peer reviewed. The authors declared no competing interests in the posted record.