SWIFT-Cell: Segmentation Without Iterative Flow Tracking

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

cs.CVOct 21, 2026

SWIFT-Cell: Segmentation Without Iterative Flow Tracking

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

Organizations: Turocrates.ai, Mumbai, India · Department of Pathology, ESIC Medical College and Hospital, Kalaburagi, Karnataka, India

Abstract

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.

Figures & tables

Method (hardware)	Time (s)	F1 vs. CPSAM	DAB +
	H&E / Ki-67 / CD3	H&E / Ki-67 / CD3	area
CPSAM 0.30 µm/px (GPU)	86 / 303 / 58	1 / 1 / 1	1.00
CPSAM 0.5 µm/px (GPU)	34 / 131 / 23	.93 / .95 / .84	1.00
QuPath (4 CPU thr.)	10.2 / 26.6 / 5.1	.68 / .54 / .61	0.68
InstanSeg [ 4 ] (GPU fp16)	1.73 / 5.95 / 1.19	.87 / .84 / .76	0.85
InstanSeg (4 CPU thr.)	78 / 199 / 49	.88 / .84 / .76	0.85

Table 1: Evaluation regions on development slides. Times: one session ( Sec. 3 ). DAB + area: median area ratio to CPSAM, matched DAB-positive Ki-67 nuclei. InstanSeg: released brightfield_nuclei v0.1.1 at 0.5 µm/px, antialiased, our tiling; GPU: its TorchScript decoder, CPU: ours. SWIFT GPU (fp16, seed dilation 5 px) and CPU (int8, 3 px) are two deployment configurations, not a quantisation comparison.

Set	Stain, tissue; reference	Patients	Nuclei	Metric	SWIFT	CPSAM	InstanSeg	StarDist	QuPath	$\Delta$ vs. CPSAM [95% CI]
PanNuke fold 1 [ 16 ]	H&E, 19 tissues; masks	no ids	63,218	PQ	.593	.627	.589	.577	.417	$-$ .033 [ $-$ .036, $-$ .030]
NuCLS [ 17 ]	H&E, breast; polygons	122	19,175	PQ	.546	.540	.524	.527	.413	$+$ .006 [.000, .011]
MoNuSeg test [ 18 ]	H&E, 7 organs; masks	14	6,697	PQ	.629	.656	.657	.665	.503	$-$ .027 [ $-$ .039, $-$ .010]
BC-DeepLIIF [ 19 ]	Ki-67, breast; tumour masks	unknown (10–15 inferred)	6,240	PQ	.609	.688	.637	.496	.344 (.405)	$-$ .079 [ $-$ .089, $-$ .064]
EndoNuke [ 20 ]	ER/PR, endometrium; points	12	210,419	F1	.882	.896	.857	.684	.803	$-$ .014 [ $-$ .019, $-$ .005]
IGNITE, 4 readers [ 14 ]	PD-L1, lung; points	10	9,676	F1	.826	.851	n.m.	n.m.	.657	$-$ .025 [ $-$ .062, $-$ .005]

Table 2: Human ground truth across stains, with patient counts where known. Pooled binary PQ [ 21 ] (IoU

{>}0.5

) for masks, detection F1 or recall (

{\leq}4

µm, 1:1) for points; no tuning on these sets. SWIFT: GPU fp16 (int8 within 0.004). QuPath in parentheses: OD-sum mode (threshold 0.1, fixed on development data). BC-DeepLIIF: 41 fields; predictions

{\geq}50

% outside all masks dropped. EndoNuke: 1,472 of 1,740 tiles by non-pathologists. IGNITE: 32 test ROIs, reader mean (inter-reader F1 0.867). Ki67-IH12: recall of 212 pathologist-outlined nuclei (centroids). InstanSeg: Table 1 settings (CPU on PanNuke and NuCLS; released pipeline within 0.007 except NuCLS, .475). CIs: paired bootstrap over patients, PanNuke patches or BC-DeepLIIF inferred cases. n.m.: not measured.

Decoder	F1 vs. CPSAM	PQ	Flips (%)
	H&E/Ki-67/CD3	PanNuke/NuCLS	H&E/Ki-67/CD3
Offsets, output	.886/.879/.728	.593/.546	6.0/25.5/7.2
Offsets, 0.5 µm	.886/.876/.727	.587/.544	20.2/34.8/21.5
Heatmap WS, output	.882/.864/.726	.590/.544	8.0/36.3/8.4
Nearest seed	.872/.842/.719	.571/.535	18.5/53.3/17.8

Table 3: Decoder ablation: final model, GPU, same seeds and foreground (bold: SWIFT). Output / 0.5 µm: labels assigned on the native / working grid. WS: watershed. Flips: matched cells whose circularity-filter decision (

{\geq}0.7

, shapes on one 0.30 µm/px grid) differs from CPSAM’s.

Figure 1: Ki-67 tuning crop (development slide). Scale bar 20 µm.

Histopathology Image Segmentation

Histopathology image segmentation identifies tissue regions, cells, and their boundaries in microscopic pathology images.

See all 53 papers

LSP-DETR: Efficient and Scalable Nuclei Segmentation in Whole-Slide Images
Jan 6, 2026Matěj Pekár, Vít Musil, Rudolf Nenutil +2Computational Pathology Image Segmentation
Masaryk University, Faculty of Informatics, Botanická 68a, Brno, 602 00, Czech Republic · Masaryk Memorial Cancer Institute, Žlutý kopec 7, Brno, 656 53, Czech Republic · BBMRI-ERIC, Neue Stiftingtalstraße 2/B/6, Graz, 8010, Austria
One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation
May 28, 2026Sanghyun Jo, Seo Jin Lee, Seohyung Hong +4Interactive Medical Image Segmentation Cross-Modality Medical Image Segmentation
1OGQ, Korea · 2Seoul National University, Korea · 3LG CNS, Korea
Revisiting foundation models for cell instance segmentation
Mar 18, 2026Anwai Archit, Constantin PapeInstance Segmentation Microscopy Image Analysis
Georg-August-University Göttingen, Institute of Computer Science · CAIMed - Lower Saxony Center for AI & Causal Methods in Medicine, Göttingen · Cluster of Excellence Multiscale Bioimaging (MBExC), Georg-August-University Göttingen

Jan 6, 2026J/K move · Enter open · S save

LSP-DETR: Efficient and Scalable Nuclei Segmentation in Whole-Slide Images

Matěj Pekár, Vít Musil, Rudolf Nenutil, Petr Holub+1

Masaryk University, Faculty of Informatics, Botanická 68a, Brno, 602 00, Czech Republic · Masaryk Memorial Cancer Institute, Žlutý kopec 7, Brno, 656 53, Czech Republic · BBMRI-ERIC, Neue Stiftingtalstraße 2/B/6, Graz, 8010, Austria

Computational Pathology Image Segmentation Histopathology Image Segmentation Instance Segmentation Detection Transformer Whole-Slide Image Analysis

Open paper AskPDF arXiv

Background and Objective: Precise and scalable instance segmentation of cell nuclei is a fundamental prerequisite for computational pathology, yet gigapixel whole-slide images (WSIs) pose significant computational challenges. While patch-based processing is standard during training, existing methods are often limited to small tile sizes during inference due to architectural bottlenecks or reliance on computationally expensive post-processing for instance separation. We introduce a faster, scalable, and end-to-end framework capable of processing large-scale image tiles while accurately modeling biologically realistic overlapping nuclei. Methods: We propose LSP-DETR (Local Star Polygon DEtection TRansformer). The model represents nuclei as star-convex polygons and employs a lightweight transformer with linear complexity, enabling the processing of high-resolution images in a single forward pass. A novel radial distance loss accommodates annotation uncertainty, allowing the segmentation of overlapping nuclei to emerge naturally without explicit overlap labels. Results: LSP-DETR achieves state-of-the-art efficiency, with an inference time of 0.45 s/mm^2, a 3.2x speedup over StarDist, the next-fastest method. On PanNuke, the model achieves competitive accuracy (67.5 bPQ), while yielding an F

_1

-score of 0.964 in polygon overlap when evaluated against consensus annotations from two expert pathologists. Furthermore, it outperforms larger models such as LKCell in generalization robustness, reaching an F1-score of 85.0 on MoNuSeg. Conclusions: LSP-DETR bridges the gap between high-fidelity segmentation and practical clinical requirements by eliminating heuristic post-processing. By providing a scalable, linear-complexity solution that naturally handles overlaps between nuclei, this framework sets a new direction for efficient high-throughput WSI analysis in digital pathology.

Abstract

Histopathology Image Segmentation

Explore similar work

LSP-DETR: Efficient and Scalable Nuclei Segmentation in Whole-Slide Images

One Click per Cell Type Suffices: Training-free Group Interaction for Cell Instance Segmentation

Revisiting foundation models for cell instance segmentation