PaperlayerView exact revision on arXiv

Source-linked AI summary

SWIFT-Cell: Segmentation Without Iterative Flow Tracking

Devansh Lalwani, Swapnil Bhat, Muralikrishna H H, Anand Reddy, Maulik Shah

arXiv:2610.33081v1cs.CV

TL;DR

Time-critical cell segmentation calls for methods that retain broad stain coverage without the latency of a generalist segmenter. This paper distils CPSAM masks into a compact U-Net and evaluates its speed and fidelity against detectors and human references. SWIFT runs faster than QuPath on the same CPU and remains close to CPSAM, with measured losses that vary across tasks.

  • Problem

    CPSAM works across stains but can be too slow for time-critical use, while fast watershed detection can miss or undersize DAB-positive nuclei.

    Source: §1 · p. 1

  • Method

    The authors distil CPSAM masks from mixed-stain tiles into a 0.63M-parameter U-Net and deploy it in GPU fp16 or CPU int8.

    Source: §1 · p. 1, §2 · p. 1

  • Results

    SWIFT runs faster than QuPath’s watershed detector on the same CPU and stays close to CPSAM in accuracy, with measured losses across evaluations.

    Source: §3 · pp. 2–3, §5 · p. 4, §4 · p. 4

  • Takeaways & Limitations

    Distilling a generalist on its own masks can bridge the deployment gap under time and hardware limits, while retaining measured accuracy losses.

    Source: §5 · p. 4

    SWIFT depends on CPSAM’s masks, so agreement with the teacher does not establish accuracy where CPSAM is unreliable.

    Source: §5 · p. 4

Abstract

from arXiv · show

Generalist segmenters such as Cellpose-SAM (CPSAM) outline cells accurately across stains but slowly; watershed detectors such as QuPath's are fast but less accurate. In time-critical use, such as interactive retrieval of cells from a user-drawn region, CPSAM at the reference settings takes 58-303 s per measured 8192 x 8192 region on an RTX 4500 Ada GPU, beyond a seconds-scale interactive target. We distil CPSAM into SWIFT-Cell (SWIFT), a 0.63M-parameter U-Net trained only on CPSAM masks of 1,202 H&E and IHC tiles. On three 8192 x 8192 development regions SWIFT takes 2.0-5.2 s on 4 CPU cores in int8, 2.5-5.1x faster than QuPath's detector on the same cores, and 0.36-0.49 s on a GPU. Against human annotations on eight H&E, IHC and ISH sets it keeps 95% of CPSAM's panoptic quality (PQ) on PanNuke (0.593 vs. 0.627) and is level on NuCLS (0.546 vs. 0.540). On Ki-67 IHC masks scored to the outline centre it is 0.205 PQ above QuPath's better mode, 0.079 below CPSAM and 0.027 below InstanSeg (CI includes 0; at the outer edge 0.008 below CPSAM); on 212 pathologist-outlined nuclei in Ki-67 fields of 12 untouched patients its recall is 0.925 (CPSAM 0.920, InstanSeg 0.901; paired difference CIs include 0); one method-blinded pathologist accepted its outlines in 96% of 100 fields (60 from two development tissue patients, 40 hematology; CPSAM 83%, QuPath 53%; tissue 100% for SWIFT and CPSAM); on ER/PR points from 12 patients its detection F1 is 0.882 (CPSAM 0.896, InstanSeg 0.857), and on 4-reader PD-L1 points 0.83 (CPSAM 0.85, readers 0.87). SWIFT is close to CPSAM in accuracy and faster than watershed detection.

1. INTRODUCTION

Generalist segmenters work across stains but can be too slow for time-critical use, while fast watershed detection can miss or undersize DAB-positive nuclei. The paper distils CPSAM masks into a small student and measures its speed and fidelity against established methods.

1. INTRODUCTION

  • Time-critical workflows need segmentation within seconds, but CPSAM can take 3.3–4.9× as long as QuPath on the measured regions.The motivating use includes retrieving cells from a user-drawn region and deployment on limited hardware.Source: §1 · p. 1
  • QuPath is fast on CPU, but its default hematoxylin-channel watershed often misses or undersizes DAB-positive nuclei.Other lightweight supervised models and teacher-output distillation exist, but the authors found no prior CPSAM distillation.Source: §1 · p. 1
  • The paper distils a trusted generalist into a 0.63M-parameter U-Net and evaluates speed, teacher fidelity, outline size, and performance against human annotations.The evaluation includes H&E and IHC annotations, zero-shot hematology, and a blinded pathologist rating.Source: §1 · p. 1

2. DISTILLING THE TEACHER

SWIFT is a compact U-Net trained on CPSAM-generated masks from a mixed H&E and IHC tile set, without human training labels. It predicts foreground, centres, and offsets for one-pass instance decoding, with GPU and CPU deployment configurations.

2. DISTILLING THE TEACHER

  • CPSAM labelled 1,202 H&E and IHC tiles, and no human labels were used to train the 0.63M-parameter student.The tiles include institutional, TCGA, and ACROBAT slides; IGNITE data were excluded from training, warm start, and calibration.Source: §2 · p. 1
  • The U-Net predicts foreground, a Gaussian centre heatmap, and pixel-to-centroid offsets, with training that up-weights the palest third of hematoxylin nuclei.The model uses a full-resolution refinement stage; that stage, the stem, and pale weighting were not ablated.Source: §2 · p. 2
  • SWIFT decodes instances from heatmap maxima inside the foreground, resamples labels to the output grid, and removes objects under five working-grid pixels.Tiles use a 48-pixel halo; GPU inference is fp16 and CPU inference uses OpenVINO int8 quantised on 64 training tiles.Source: §2 · p. 2
  • Validation tiles may overlap warm-start training data, so validation did not select the model and is not independent.The final checkpoint was used after fine-tuning on the full tile set.Source: §2 · p. 2

3. RESULTS ON TISSUE

On development regions, SWIFT is faster than QuPath and better recovers DAB-positive nuclei than QuPath’s default detector. Results against human references vary by dataset and boundary convention, with additional differences in decoder choices and test-set overlap affecting interpretation.

3. RESULTS ON TISSUE

  • The timed regions are on development slides, and the measured hardware, inferred BC-DeepLIIF patients, and unstated boundary convention constrain interpretation.The study reports no human-mask evidence for CD3 and does not measure downstream retrieval performance.Source: §3 · pp. 2–3
  • SWIFT’s int8 segmenter is 1.7–3.0× faster than QuPath on the same CPU cores; it misses 2.2% of CPSAM’s DAB-positive Ki-67 nuclei and recovers 91% at IoU≥0.5.Default QuPath misses 39% outright and outlines a further 20% at under half their area; its OD-sum mode recovers 76%.Source: §3 · p. 2
  • Decoder ablations show that offsets help assignment and output-grid labels help shape; working-grid labels keep F1 within 0.003 but change 20–35% of circularity decisions.Heatmap watershed on the output grid comes within 0.004 of offsets except on dense Ki-67.Source: §3 · p. 2, Table 3 · p. 3
  • Table 2 compares pooled binary PQ for masks and detection F1 or recall for point annotations across human-ground-truth datasets.It reports patient counts where known and specifies that there was no tuning on these sets.Source: Table 2 · p. 3
  • SWIFT’s F1 is 0.018 below CPSAM on NuCLS, while on MoNuSeg test InstanSeg and StarDist score 0.028 and 0.036 PQ above SWIFT.On MoNuSeg, those models have higher precision (0.87, 0.89 vs. 0.77) and lower recall (0.84 vs. 0.90); the test patients are outside SWIFT’s training set.Source: §3 · p. 3, Table 2 · p. 3
  • On 212 pathologist-outlined nuclei from untouched Ki-67 fields, SWIFT recall is 0.925, versus 0.920 for CPSAM and 0.901 for InstanSeg.Paired difference confidence intervals versus CPSAM and InstanSeg include zero; SWIFT’s matched outlines average 1.5× the pathologist’s area.Source: §3 · p. 3, Table 2 · p. 3
  • SWIFT is 0.205 PQ above QuPath OD-sum on BC-DeepLIIF, but 0.027 below InstanSeg at the outline centre and 0.008 below CPSAM at the outer edge.The InstanSeg difference at the outline centre has a confidence interval including zero; the boundary convention changes the ranking.Source: §3 · p. 3, Table 2 · p. 3
  • On EndoNuke ER/PR points, InstanSeg has higher precision than SWIFT (0.93 vs. 0.87) and lower recall (0.79 vs. 0.89); SWIFT recalls 86% of SiNuS HER2 dual-ISH masks.On pale ISH nuclei, SWIFT can outline signal clusters, splitting 14–16% of nuclei.Source: §3 · p. 3, Table 2 · p. 3

4. ZERO-SHOT HEMATOLOGY

Without hematology training or tuning, the tissue-configured student performs differently on smears and marrow aspirates: it splits many smear leukocytes but covers many compact marrow blobs. One blinded pathologist rated its outlines more acceptable than CPSAM’s and QuPath’s on the sampled fields.

4. ZERO-SHOT HEMATOLOGY

  • On peripheral smears, the tissue-configured student splits 70% of 84 CPSAM leukocytes, with F1 0.19 against CPSAM’s leukocyte objects.The 40 marrow counting fields came from two held-out patients; the student covered 76% of compact chromatin blobs, a proxy that ignores false detections, merges, and boundaries.Source: §4 · pp. 3–4
  • One method-blinded pathologist accepted SWIFT outlines in 96% of 100 fields, compared with 83% for CPSAM and 53% for QuPath.The fields include 60 tissue fields from two development patients and 40 hematology fields; the findings describe one reader’s judgements on these fields.Source: §4 · p. 4
  • SWIFT minus CPSAM acceptability was +13 points, with a field-level 95% CI of [+6, +21], but few-cluster intervals may under-cover.Noninferiority on a tissue population is not established.Source: §4 · p. 4

5. LIMITATIONS AND CONCLUSION

SWIFT is faster than QuPath’s watershed detector on the same CPU and remains close to CPSAM in accuracy, with measured losses. Its evidence is bounded by teacher reliability and evaluation settings that do not establish performance on a T4 or downstream retrieval.

5. LIMITATIONS AND CONCLUSION

  • Agreement with CPSAM does not establish accuracy where its masks are unreliable; SWIFT’s teacher-PQ deficits are 0.033 on PanNuke, 0.027 on MoNuSeg, and 0.079/0.008 on BC-DeepLIIF at the outline centre/outer edge.SWIFT also changes CPSAM’s circularity-filter decision for 6–26% of matched cells.Source: §5 · p. 4
  • The timed regions were on development slides, CD3 lacks human-mask evidence, and times used 4 EPYC cores and an RTX 4500 Ada rather than a T4; downstream retrieval was not measured.BC-DeepLIIF patients were inferred, and its unstated boundary convention changes the InstanSeg ranking.Source: §5 · p. 4
  • The 0.63M-parameter student runs faster than QuPath’s watershed detector on the same CPU while staying close to the teacher in accuracy, with measured losses.The authors conclude that distilling a research generalist on its own masks can bridge the gap to deployment under time and hardware limits.Source: §5 · p. 4
Loading 2610.33081v1…