IMVS: Interactive Medical Volume Segmentation with Test-Time Adaptation - A New Method for Annotating Radiology Datasets
Abhilaksh Singh Reen, Kushal Borkar, Ritvik Mahapatra
cs.CV
Quick verdict
Claims 1.9–14.4× faster medical-volume annotation via online-adapted 2D adapter + frozen tracker; promising, honestly scoped, but n=9 user study and stats are thin.
Summary
IMVS targets the bottleneck of annotating 3D radiology volumes, where slice-wise interactive tools force repeated corrections and fixed 3D/video propagators drift on hard anatomy. Rather than a new segmentation primitive, it composes a closed annotation loop: a lightweight 2D Slice Mask Adapter (UNet++/DeepLabV3+/TransUNet) fine-tuned online from user scribbles, a frozen ViT-B Volume Mask Tracker (YouTube-VOS-pretrained, medically fine-tuned) that propagates accepted masks for up to K=17 slices, and a soft teacher-student BN alignment that limits forgetting. Training uses only three datasets (BraTS, LiTS, MSD Pancreas) with simulated scribbles; five datasets are evaluated zero-shot. In a nine-resident user study with a shared acceptance criterion (DSC 0.885), IMVS matched baseline quality at far lower cost: 14.4x faster than copy-based manual annotation (22.3x vs naive manual), 4.6x vs slice-wise and 1.9x vs 3D interactive methods, fastest on 5 of 8 datasets and lowest VRAM (3.23 GB). It reaches DSC >= 0.90 with the fewest interactions everywhere (7.1 mean vs 10.1 for MedSAM2). A Wilcoxon test vs MedSAM2 shows significance only on challenging targets (p = 0.02; p = 0.26 on well-delineated organs, where MedSAM2 stays competitive). Ablations attribute gains mainly to online SMA adaptation (18 to 4 interactions); freezing the VMT avoids 6x VRAM and 14x latency. Limitations include single-target-per-pass operation, reduced benefit on tortuous structures, and a small user study.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Composing an online-adapted 2D Slice Mask Adapter with a frozen Volume Mask Tracker amortizes each user correction across slices, cutting annotation time 14.4x vs copy-based manual and 1.9x vs 3D interactive methods (Table 1).
- IMVS reaches DSC >= 0.90 with the fewest interactions on every evaluated dataset (7.1 mean vs 10.1 for MedSAM2, Table 2).
- Freezing the VMT is essential: adapting it yields no accuracy gain but 6x VRAM and 14x latency (Table 3).
- The advantage is statistically significant on challenging targets (LiTS, Pancreas, HepaticVessel, BraTS; p=0.02) but not on well-delineated organs (p=0.26), where MedSAM2 remains competitive.
Strengths
Honest, tightly scoped claims
Abstract and §4.2 explicitly concede MedSAM2/ScribblePrompt stay competitive or stronger on well-delineated organs, report the non-significant p=0.26, and present 14.4x (not 22.3x) as the fair comparison against copy-based manual annotation.
Broad, well-structured empirical grounding
Eight public CT/MRI datasets with a clean 3-train/5-zero-shot split (§3.5), seven interactive baselines, and three complementary views: time + interactions (Table 1), interactions at matched quality DSC>=0.90 (Table 2), and component ablation with VRAM/latency (Table 3).
Component ablation isolates the mechanism
Strong reproducibility posture for a method paper
Weaknesses
Thin human-study evidence under the headline claim
Minimal statistical treatment
Simulator/human protocol blur
Internal inconsistencies in reported numbers
Practical scope limits
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
IMVS: Interactive Medical Volume Segmentation with Test-Time Adaptation — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is IMVS and how much faster is it than manual annotation?
IMVS is an interactive annotation framework for 3D medical (CT/MRI) volumes. A user scribbles corrections on one slice; a lightweight 2D network learns from them online while a frozen tracker propagates the corrected mask to neighboring slices. Across eight public datasets in a nine-resident study, it was 14.4x faster than a proficient copy-based manual workflow, 22.3x faster than naive manual annotation, 4.6x faster than slice-wise interactive tools, and 1.9x faster than 3D interactive methods.
How does IMVS differ from MedSAM2?
MedSAM2 propagates a prompt through a volume with a fixed pretrained model that never adapts to the target scan, so it can drift on low-contrast or pathological structures. IMVS additionally fine-tunes a small 2D adapter online from each user correction, so the system stops repeating the same errors. On well-delineated organs MedSAM2 is equally fast or faster; IMVS's advantage concentrates on challenging targets (tumors, vessels) and it needs fewer interactions to reach DSC >= 0.90 (7.1 vs 10.1 on average).
Is the code and data available for IMVS?
The authors state code and a demo video are at github.com/AbhilakshSinghReen/imvs. All eight evaluation datasets are public (CHAOS, AMOS, MSD Prostate/HepaticVessel/Pancreas, LiTS, BraTS), and training scribbles are simulated from ground truth, so the algorithmic experiments are in principle reproducible — though the human timing study itself cannot be rerun exactly.
What are the main limitations of IMVS?
It segments one target structure per pass, so multi-label volumes need sequential sessions. It helps least on tortuous or discontinuous structures such as hepatic vessels, where the tracker holds only about 4-6 slices before needing correction. The efficiency claims rest on a nine-resident user study with limited statistical treatment, and the paper is a workshop-track manuscript not yet through peer review.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are not analyzable from the extracted text; Fig. 1's architecture is judged from its caption only.
- Table column alignment in the extraction is partially garbled, so per-column readings were cross-checked against in-text statements (they are consistent for the key claims).
- The GitHub repository contents were not inspected; code availability is as stated by the authors.
- This is arXiv v1 of a workshop-accepted manuscript, explicitly prior to peer review, as the authors note in the footnote.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.