Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Matěj Pekár, Vít Musil, Rudolf Nenutil, Petr Holub, Tomáš Brázdil

TL;DR

The paper tackles nuclei instance segmentation in gigapixel whole-slide images, where data variability and image size challenge patch-based approaches. It introduces LSP-DETR, an end-to-end DETR-based framework that models nuclei as star-convex polygons described by a center and 64 radial distances, refined via a lightweight transformer. A radial distance loss defines per-ray bounds for foreground points, allowing overlapping nuclei segmentation without explicit overlap annotations or post-processing, and enabling linear complexity through Sliding Tile Attention. On PanNuke and MoNuSeg, LSP-DETR achieves state-of-the-art speed and strong generalization, with an inference speed of 0.46 s/mm^2 and linear complexity, demonstrating practical applicability to high-throughput WSI pipelines.

Abstract

Precise and scalable instance segmentation of cell nuclei is essential for computational pathology, yet gigapixel Whole-Slide Images pose major computational challenges. Existing approaches rely on patch-based processing and costly post-processing for instance separation, sacrificing context and efficiency. We introduce LSP-DETR (Local Star Polygon DEtection TRansformer), a fully end-to-end framework that uses a lightweight transformer with linear complexity to process substantially larger images without additional computational cost. Nuclei are represented as star-convex polygons, and a novel radial distance loss function allows the segmentation of overlapping nuclei to emerge naturally, without requiring explicit overlap annotations or handcrafted post-processing. Evaluations on PanNuke and MoNuSeg show strong generalization across tissues and state-of-the-art efficiency, with LSP-DETR being over five times faster than the next-fastest leading method. Code and models are available at https://github.com/RationAI/lsp-detr.

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

TL;DR

The paper tackles nuclei instance segmentation in gigapixel whole-slide images, where data variability and image size challenge patch-based approaches. It introduces LSP-DETR, an end-to-end DETR-based framework that models nuclei as star-convex polygons described by a center and 64 radial distances, refined via a lightweight transformer. A radial distance loss defines per-ray bounds for foreground points, allowing overlapping nuclei segmentation without explicit overlap annotations or post-processing, and enabling linear complexity through Sliding Tile Attention. On PanNuke and MoNuSeg, LSP-DETR achieves state-of-the-art speed and strong generalization, with an inference speed of 0.46 s/mm^2 and linear complexity, demonstrating practical applicability to high-throughput WSI pipelines.

Abstract

Precise and scalable instance segmentation of cell nuclei is essential for computational pathology, yet gigapixel Whole-Slide Images pose major computational challenges. Existing approaches rely on patch-based processing and costly post-processing for instance separation, sacrificing context and efficiency. We introduce LSP-DETR (Local Star Polygon DEtection TRansformer), a fully end-to-end framework that uses a lightweight transformer with linear complexity to process substantially larger images without additional computational cost. Nuclei are represented as star-convex polygons, and a novel radial distance loss function allows the segmentation of overlapping nuclei to emerge naturally, without requiring explicit overlap annotations or handcrafted post-processing. Evaluations on PanNuke and MoNuSeg show strong generalization across tissues and state-of-the-art efficiency, with LSP-DETR being over five times faster than the next-fastest leading method. Code and models are available at https://github.com/RationAI/lsp-detr.
Paper Structure (11 sections, 4 figures, 1 table)

This paper contains 11 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Large data variability
  4. Large image size
  5. State of the Art
  6. Contribution: LSP-DETR
  7. Related Work
  8. Two-Stage Detectors
  9. One-Stage Detectors with Post-processing
  10. DETR-Based Approaches
  11. LSP-DETR Model

Figures (4)

  • Figure 1: Nuclei instance segmentation on a whole-slide image. The goal is to segment all individual cell nuclei. The algorithm must be scalable and robust to various tissue types.
  • Figure 2: Ambiguous boundaries and overlaps pose one of the challenges in nuclei segmentation.
  • Figure 3: Illustration of the LSP-DETR training process.
  • Figure :