Skip to content
Log in
Sponsored Artificial Intelligence: A Modern Approach, Global Edition Artificial Intelligence: A Modern Approach, Global Edition

Author

Kundan Thind

View on Semantic Scholar

2 papers indexed here

We haven’t gathered this author’s papers yet. Follow them and we’ll fetch their work.

Not the right person? Other researchers publish under this name.

Open access Sep 2026

Cross-fraction prior learning for scalable organ-at-risk segmentation in abdominal MR-guided radiotherapy.

BACKGROUND Manual organ-at-risk (OAR) delineation takes 20-40 min per case, a major bottleneck within the 50-90 min treatment window of abdominal MR-guided adaptive radiotherapy (MRgRT). Most deep learning systems adopt single-fraction approaches that discard valuable temporal context from prior treatment fractions. PURPOSE This study develops AdaptSeg, a scalable framework leveraging cross-fraction anatomical priors to substantially improve OAR segmentation without per-patient retraining. METHODS We implemented a dual-path neural architecture conditioning current fraction segmentation on paired image-mask information from supporting fractions. AdaptSeg was instantiated with convolutional (3D UNet) and transformer-based (SwinUNETR) backbones. Evaluation used 104 pancreatic cancer patients across 520 treatment fractions for four abdominal organs (colon, duodenum, small bowel, stomach), with patient-level splitting: 72 training, 10 validation, 22 test patients. Performance metrics included Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD); paired comparisons used two-sided Wilcoxon signed-rank tests with Benjamini-Hochberg correction, and 95% bootstrap confidence intervals for the means. RESULTS Cross-fraction priors improved segmentation performance for both tested backbones. The 3D UNet achieved 87.22% mean DSC versus 83.78% baseline, while SwinUNETR reached 85.19% versus 82.49% baseline. For highly deformable organs, improvements included up to 7.0 percentage point DSC gains (small bowel: 77.8% to 84.8%, p < 0.001 ) and 62% boundary error reduction (colon HD95: 23.13 to 8.74 mm, p < 0.001 ). Compared to nine state-of-the-art methods, AdaptSeg achieved the best overall performance with substantial improvements in mean DSC (4.1%), HD95 (43%), and ASSD (39%) over the strongest baseline. All variants maintained computationally feasible inference under 1.6 s per case. Temporal prior selection showed a backbone-dependent preference: the CNN favored sequential priors, and the transformer additionally benefited from randomized support sampling during training (sequential support is used at inference). CONCLUSIONS Cross-fraction anatomical priors improved OAR segmentation for both tested backbone families, indicating that temporal context is an underutilized resource in fractionated radiotherapy. AdaptSeg provides a scalable, computationally feasible framework for accelerating MRgRT workflows without per-patient adaptation, with sub-1.6 s inference compatible with the time constraints of online adaptive treatment.

Chengyin Li, D. Rusu, Rafi Ibn Sultan et al. · 0 citations
Preprint Aug 2026

Model Confidence Under Answer-Preserving Attacks: An Informativeness-Manipulability Frontier

Deployed vision-language systems often gate their answers on confidence, making confidence robustness relevant to oversight. We study confidence readouts under white-box, image-only attacks constrained to preserve the generated answer byte-identically. Under a reachability assumption, an unmovable readout cannot outperform the answer-string accuracy prior, whose pooled value is 0.617. Independently of that assumption, a uniform amplitude certificate below a measurable threshold guarantees adversarial discrimination above the same floor. Across four vision-language models, three visual question answering benchmarks, five deployed confidence channels and two defense estimators, direct or surrogate-aimed attacks produce itemwise feasible perturbations that refute this uniform certificate in all 84 estimator-by-cell combinations. Coordinated correctness-label-aware attacks drive adversarial discrimination to or below the answer-string floor in all sixty deployed-channel cells, including all fifty-nine that begin above it. Hidden-state interventions and an open-ended text-model activation-space replication show that comparable confidence movement can be induced at the representation level rather than only through adversarial images. None of four tested defense families establishes a robust alternative under the specific evaluation applied to it. In a confidence-gated simulation, a coordinated token-probability attack transferred to a hidden-state gate causes up to 84.8% of previously rejected wrong answers to become accepted. After reweighting to each benchmark's natural correctness prevalence, accepted accuracy falls below the no-gate baseline in eight of twelve cells under transfer and all twelve under a direct gate-aimed attack. Under the studied threat model and budget, confidence is therefore an integrity-sensitive rather than intrinsically robust oversight signal.

Mohammad M. Ghassemi, Ivan Brugere, Simerjot Kaur et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.