• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Imperfect Bounding Boxes Degrade Prompt-Based Medical Image Segmentation Performance

by
November 21, 2026
in Technology
Reading Time: 7 mins read
0
Imperfect Bounding Boxes Degrade Prompt-Based Medical Image Segmentation Performance

Imperfect Bounding Boxes Degrade Prompt-Based Medical Image Segmentation Performance

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Standard evaluations of prompt-based vision models such as the Segment Anything Model and its specialized medical adaptation MedSAM frequently rely on idealized spatial prompts derived directly from reference annotations, a practice that risks substantially overestimating their reliability in real-world clinical environments. In clinical workflows, spatial prompts such as bounding boxes are rarely flawless, as they are routinely affected by user variability, localization uncertainty, and the inherent inaccuracies of automated detection systems. To systematically evaluate how imperfect spatial guidance affects segmentation performance, researchers Nicolas Martin, Jean-Pierre Chevallet, and Philippe Mulhem assessed the ViT-B variants of the original Segment Anything Model and MedSAM across three distinct two-dimensional medical imaging datasets covering skin lesion dermoscopy, grayscale spheroid microscopy, and bone tumor radiography. Their investigation demonstrated that both architectures remain highly sensitive to spatial prompt perturbations, revealing that the geometric nature of an error strongly dictates the degree of performance deterioration.

The study specifically addressed the question of how prompt degradation influences segmentation accuracy by subjecting ideal bounding boxes to controlled geometric modifications across predefined magnitude intervals. In prompt-based segmentation, the model does not predict segmentation masks exclusively from raw pixel content; rather, an image encoder, prompt encoder, and mask decoder jointly process image representations and prompt embeddings to direct predictions toward a designated region of interest. While standard benchmark protocols commonly extract the minimum axis-aligned bounding box enclosing ground-truth masks to simulate inference, practical clinical systems cannot guarantee optimal spatial framing. To examine whether medical domain adaptation inherently confers robustness against such deviations, the authors subjected the models to five discrete perturbation levels relative to the ideal box dimensions: greater than zero to five percent, five to ten percent, ten to fifteen percent, fifteen to twenty percent, and twenty to twenty-five percent. Across each level, perturbation parameters were randomly sampled to mimic the localization variability encountered during routine clinical deployment.

To avoid data overlap with images potentially encountered during the pretraining or validation stages of the evaluated architectures, the research team restricted their evaluation to three independent two-dimensional benchmarks released after model development or excluded from MedSAM training. The first corpus comprised the University of Waterloo skin cancer database, containing one hundred and eighty RGB dermoscopic images paired with binary segmentation masks for melanoma lesions. The second corpus was derived from the Spheroid Light Microscopy Image Atlas, known as SLiMIA, which features grayscale brightfield images of multicellular spheroids characterized by substantial morphological heterogeneity, boundary irregularity, and varying cell density; to limit acquisition confounders while preserving biological variability across distinct cell lines, a standardized subset of fifty-one images captured on an Axiovert two-hundred M microscope was analyzed. The third benchmark consisted of the Bone Tumor X-ray Radiograph dataset, designated BTXRD, from which a homogeneous cohort of one thousand six hundred and twenty-seven radiographs featuring a single annotated primary bone tumor was selected.

The experimental protocol categorized spatial prompt conditions into three primary frameworks: ideal prompts, heuristic prompts, and perturbed bounding boxes. Ideal prompts established the reference performance ceiling using minimum axis-aligned bounding rectangles generated directly from reference masks. In contrast, heuristic prompting strategies were introduced to determine whether generic or collection-level spatial priors could adequately substitute for localized guidance in the absence of exact object localization. These heuristic configurations included an agnostic centered bounding box covering fifty percent of the image height and width, a full-image bounding box covering the entire field of view to remove all localized spatial guidance, and a most probable location box computed from the mean coordinate boundaries of the training sets. For the perturbed framework, controlled transformations were executed across three distinct geometric operations: spatial shift, which translated the entire box along horizontal and vertical axes while preserving bounding box dimensions; shrinkage, which reduced box dimensions symmetrically around the center to partially exclude the target; and expansion, which symmetrically enlarged the box beyond target margins.

To quantify model behavior beyond traditional overlap measures, the researchers formulated a multidimensional evaluation framework consisting of the Dice similarity coefficient, Normalized Correction Effort, and Relative Performance Degradation. The Dice similarity coefficient served as the primary measure of spatial overlap between predicted masks and ground-truth reference masks, ranging from zero to one. To capture the clinical burden of manually refining imperfect predictions, the authors introduced Normalized Correction Effort, defined as the ratio of false positive and false negative pixels to the total union of predicted and ground-truth masks. This formulation reflects the pixel-wise symmetric difference, providing a normalized proxy for the residual interaction required to convert a model prediction into an acceptable clinical annotation. Additionally, Relative Performance Degradation was defined as the percentage drop in Dice score relative to the score achieved under ideal bounding box conditions, allowing direct comparisons of prompt sensitivity across disparate datasets, perturbation scales, and network variants independent of baseline performance.

Statistical evaluation was conducted using linear mixed-effects models implemented in R with the lme4 library to account for repeated measures and hierarchical dependencies across the experimental configurations. Perturbation type, perturbation magnitude, and model architecture were modeled as fixed effects, whereas individual image identity and repeated stochastic runs were incorporated as random effects. Across all evaluated conditions, each image was tested across five seeded runs to capture stochastic variability, and the random effect associated with repeated runs consistently proved negligible, demonstrating high stability in the perturbation generation mechanism. Linear mixed-effects analyses confirmed significant main effects and interaction effects between perturbation type and perturbation magnitude across Dice, Normalized Correction Effort, and Relative Performance Degradation metrics, with post-hoc pairwise Tukey-adjusted comparisons establishing statistically significant differences between all perturbation geometries at significance thresholds below p equals zero point zero zero zero one.

On the Waterloo skin lesion dataset, segmentation performance exhibited a steep progressive decline as perturbation magnitude increased, with the degradation profile heavily governed by perturbation geometry. Shrinking perturbations caused severe drops in segmentation accuracy for both architectures: MedSAM Dice scores collapsed from eighty-seven point four plus or minus seven point zero percent at zero to five percent shrinkage down to thirty-three point eight plus or minus six point four percent at twenty to twenty-five percent shrinkage. The original Segment Anything Model similarly fell from ninety point nine plus or minus eight point zero percent to forty-eight point two plus or minus eleven point six percent. Parallel surges occurred in Normalized Correction Effort and Relative Performance Degradation, demonstrating that excluding portions of the lesion substantially elevated manual correction requirements. Conversely, bounding box expansion remained remarkably well tolerated, with both models maintaining Dice overlap scores above eighty percent even when bounding boxes were expanded by twenty to twenty-five percent beyond the ground truth.

Experimental outcomes on the SLiMIA spheroid microscopy dataset reinforced this pronounced geometric asymmetry, while also highlighting unexpected performance divergences between the generalist and medically adapted models. Shrinking perturbations again produced the most detrimental impact, causing MedSAM Dice scores to plummet from ninety-four point two plus or minus four point three percent at minor shrinkage to thirty point nine plus or minus six point eight percent at twenty to twenty-five percent shrinkage, with Relative Performance Degradation exceeding sixty-five percent and Normalized Correction Effort surpassing eighty percent. Under identical shrinkage conditions, the general-domain Segment Anything Model demonstrated noticeably higher robustness, declining from ninety-eight point one plus or minus one point zero percent to fifty-eight point one plus or minus seven point zero percent. Significant three-way interactions between model type, perturbation geometry, and perturbation magnitude confirmed that the original Segment Anything Model maintained superior stability under several shrinking and translation conditions within this grayscale microscopy domain.

Evaluation on the bone tumor radiograph dataset further emphasized how anatomical imaging domains influence model vulnerability and prompt sensitivity. Unlike the dermoscopy and microscopy datasets where shrinkage consistently proved more damaging than translation, the BTXRD radiographs revealed distinct structural degradation patterns: for the Segment Anything Model, the estimated Dice degradation slope was negative one point zero five nine for shrinking and negative one point one one six for shifting, indicating comparable vulnerability to spatial translation. For MedSAM, the Dice degradation slope was negative one point six one two for shrinking compared to negative one point one three six for shifting, with absolute Dice decreasing from sixty point three plus or minus eighteen point four percent to twenty-eight point five plus or minus thirteen point two percent under shrinkage, and from sixty-two point four plus or minus eighteen point eight percent to forty point zero plus or minus eighteen point four percent under shift. Bounding box expansion again proved markedly less disruptive, allowing MedSAM to sustain Dice scores above sixty percent across all evaluated expansion intervals.

The evaluation of collection-level heuristic prompting strategies demonstrated that prompt-conditioned architectures cannot function effectively without localized spatial priors. Across all three imaging datasets, full-image prompts produced severe performance breakdowns, yielding Dice scores of approximately fifteen percent on Waterloo, under ten percent on SLiMIA, and three point six percent on BTXRD, accompanied by Normalized Correction Effort values exceeding ninety to ninety-eight percent. Similarly, dataset-level most probable location prompts and centered agnostic boxes yielded low Dice scores that remained below ten percent on the radiograph collection despite the general central positioning of pathological features. These findings emphasize that while prompt-based foundation models possess strong segmentation capabilities when guided by localized cues, they lack autonomous object localization capacity when spatial conditioning is diluted to broad regional heuristics.

Synthesizing these findings, the authors observed that prompt-conditioned models exhibit a fundamental asymmetric tolerance: they are far more sensitive to the exclusion of relevant target regions than to the inclusion of surrounding background context. When a bounding box expands, the target structure remains entirely encapsulated, preserving critical spatial priors despite the introduction of peripheral ambiguity; conversely, when a bounding box shrinks or shifts outside object boundaries, partial exclusion of the target structure rapidly induces segmentation failure. Crucially, domain-specific pretraining in MedSAM did not systematically confer increased robustness against imperfect prompts compared to the base Segment Anything Model. These results suggest that prompt generation systems and automated detection pipelines should deliberately incorporate safety margins to prioritize complete target coverage over tight spatial bounding, ensuring that imperfect spatial prompts expand rather than truncate anatomical boundaries.

The authors noted several methodological boundaries governing the scope and interpretation of their conclusions. The primary experimental framework examined isolated geometric transformations along individual axes rather than compound or irregular spatial distortions that could arise during complex clinical interactions. Furthermore, the analysis was deliberately confined to two-dimensional image-wise segmentation across the ViT-B implementations of the Segment Anything Model and MedSAM, precluding direct extrapolation to volumetric imaging, sequential video tracking, or alternative medical adaptations such as SAM-Med2D and MedSAM2 without dedicated empirical adaptation. Additionally, the microscopy evaluation utilized a restricted subset of fifty-one SLiMIA images acquired under uniform laboratory parameters to minimize technical confounders, thereby limiting biological and instrumental heterogeneity. Consequently, the researchers emphasized that segmentation accuracy and prompt robustness constitute distinct operational properties, concluding that future clinical deployments must incorporate robustness-oriented benchmarking, adaptive prompt refinement, and uncertainty-aware mechanisms to handle imperfect real-world spatial guidance.

Subject of Research: Robustness of prompt-based segmentation models under spatial bounding box perturbations in medical imaging

Article Title: On the robustness of prompt-based medical image segmentation: the impact of bounding box quality

Article References: Martin, N., Chevallet, J.-P., & Mulhem, P. (2026). On the robustness of prompt-based medical image segmentation: the impact of bounding box quality. Multimedia Tools and Applications, 85(11), Article 833. https://doi.org/10.1007/s11042-026-21982-4

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21982-4

Keywords: Medical Image Segmentation, Segment Anything Model, MedSAM, Bounding Box Quality, Spatial Perturbations, Waterloo Dataset, SLiMIA Dataset, BTXRD Dataset, Dice Similarity Coefficient, Normalized Correction Effort, Relative Performance Degradation, Linear Mixed-Effects Models, Skin Lesion Imaging, Spheroid Microscopy, Bone Tumor Radiographs, Heuristic Prompting, Domain Adaptation

News Source: Denise Maddox. (November 21, 2026). Imperfect Bounding Boxes Degrade Prompt-Based Medical Image Segmentation Performance. Scienmag.

Tags: Bone Tumor RadiographsBounding Box QualityBTXRD DatasetDice similarity coefficientdomain adaptationHeuristic Promptinglinear mixed-effects modelsmedical image segmentationMedSAMNormalized Correction EffortRelative Performance DegradationSegment Anything ModelSkin Lesion ImagingSLiMIA DatasetSpatial PerturbationsSpheroid MicroscopyWaterloo Dataset
Share12Tweet7Share2ShareShareShare1

Related Posts

Neural Network With Sliding Data Window Enhances Autonomous Ship Path Following

Neural Network With Sliding Data Window Enhances Autonomous Ship Path Following

November 22, 2026
Crayfish Optimization Paired with Gaussian Regression Accurately Predicts Mine Blasting Fragmentation

Crayfish Optimization Paired with Gaussian Regression Accurately Predicts Mine Blasting Fragmentation

November 22, 2026

Radio Tags Alter Flight Kinematics in Neotropical Butterflies Below Standard Weight Limits

November 22, 2026

Strut-and-Tie Model Accurately Predicts Flexural Capacity of Prefabricated Bridge Wet Joints

November 22, 2026

POPULAR NEWS

  • Urinary Bacteria Produce Male Hormones Through Newly Identified Microbial Enzyme Pathway

    Urinary Bacteria Produce Male Hormones Through Newly Identified Microbial Enzyme Pathway

    29 shares
    Share 12 Tweet 7
  • Tislelizumab Combined With Chemotherapy Extends Long-Term Survival in Advanced Esophageal Cancer

    29 shares
    Share 12 Tweet 7
  • Global Survey Reveals Marked Variation in Anticoagulation Practices During Adult ECPR

    29 shares
    Share 12 Tweet 7
  • Mutant p53 Reprograms AP-1 Complexes to Drive Cancer Cell Migration

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Infant Gut Metabolites Link Cesarean Delivery to Elevated Childhood Asthma Risk

Neural Network With Sliding Data Window Enhances Autonomous Ship Path Following

Crayfish Optimization Paired with Gaussian Regression Accurately Predicts Mine Blasting Fragmentation

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.