detect and segment cells

In digital pathology quality control, manual verification of tissue fragment counts is time-consuming, highly subjective, and exhibits substantial inter-observer variability (inter-expert agreement: 82–88%), thereby compromising diagnostic reliability. To address this, we propose the first automated fragment counting method integrating YOLOv9 and Vision Transformer architectures, augmented with medical image-specific preprocessing and postprocessing optimizations. Evaluated on real-world clinical whole-slide images, our method achieves an 86% counting accuracy—within the observed range of inter-expert variability. By substantially reducing subjective bias and demonstrating robust performance on routine clinical data, the approach exhibits strong feasibility for clinical deployment. This work establishes a generalizable technical paradigm for end-to-end automated quality control in digital pathology workflows.

To address challenges in microscopic imaging—including dense small-object arrangements, large scale variations, and scarce annotations—this paper proposes an intelligent analysis system tailored for microscopy. Methodologically, we design a dedicated instance segmentation network enabling robust separation of up to thousands of tightly packed objects; develop a human-in-the-loop data engine integrating real acquisition, controllable synthetic generation, and closed-loop annotation to alleviate labeling bottlenecks; and incorporate multi-source data fusion with an OCR-based scale bar recognition module for automatic scale calibration. Our key contributions are the first microscopy-specific instance segmentation paradigm and a scalable, synergistic data generation framework. Evaluated on diverse biological and materials microscopy datasets, the system achieves a mean average precision (mAP) of 92.3% for segmentation and 98.7% accuracy for scale bar recognition. It has been successfully deployed and is operating stably across three interdisciplinary experimental platforms.

This study addresses efficient cell detection in resource-constrained digital cytological pathology imaging. We systematically evaluate StarDist, Cellpose, SAM2, and fully convolutional regression networks (FCRN/IFCRN) on Pap-stained whole-slide images, proposing a distance-based cell localization metric and an improved FCRN (IFCRN) featuring a lightweight architecture and targeted data augmentation. Evaluated on the CNSeg and oral cancer datasets, IFCRN achieves superior performance: mAP improvements of 3.2–5.7% over state-of-the-art instance segmentation methods, 42–68% reduction in GPU memory consumption, and inference latency under 10 ms per cell. These results demonstrate that IFCRN delivers high detection accuracy with significantly lower computational overhead, offering a practical solution for real-time clinical decision support in cytology.

Existing cell segmentation methods either rely on manual parameter tuning and strong shape assumptions or require large amounts of annotated data, limiting generalizability to novel cell types, noisy images, and those with severe intensity inhomogeneity. This paper introduces the first unsupervised cell segmentation framework based on fast Gaussian processes—requiring no labels, zero parameter tuning, and making no prior assumptions about cell morphology. Our key contributions are: (1) modeling the joint spatial-intensity distribution of pixels via a scalable Gaussian process; (2) designing an adaptive, robust thresholding criterion that suppresses noise while preserving weak boundaries; and (3) integrating an optimized watershed algorithm for accurate separation of touching cells. Extensive evaluation on synthetic benchmarks and large-scale real microscopy datasets demonstrates that our method significantly outperforms state-of-the-art supervised and traditional unsupervised approaches in segmentation accuracy, robustness to noise and intensity variation, and computational scalability.

To address the limitations in microscopy cell counting—namely, model complexity and the entanglement of localization and counting tasks, which jointly hinder accuracy and efficiency—this paper proposes a decoupled dual-network framework. The framework separates counting and localization: a lightweight Counter network extracts global features to generate a coarse density map and outputs an accurate total cell count; a Locator network, conditioned on both the original image and the coarse map, reconstructs a high-resolution density map for precise single-cell localization. To alleviate optimization difficulties inherent in direct high-resolution density map regression, we introduce a novel cross-regional global message-passing module. The architecture employs a compact two-branch CNN with intermediate-layer feature fusion and conditional density map reconstruction. Evaluated on four standard benchmarks, our method achieves state-of-the-art performance, significantly reducing average counting error. The source code is publicly available.

This study addresses the challenges in classifying acute lymphoblastic leukemia (ALL) cells from blood smears, where low cytoplasmic contrast and high morphological variability render conventional membrane-based segmentation methods ineffective and limit the generalization of existing deep learning models. To overcome these limitations, the authors propose PRISM, a novel approach that eschews precise cytoplasmic segmentation and instead constructs adaptive concentric rings around the nucleus. This framework integrates color features with gray-level co-occurrence matrix–derived texture information and employs a calibrated stacked ensemble classifier for robust discrimination. By eliminating reliance on cell boundary delineation, PRISM demonstrates strong robustness and generalizability across diverse staining protocols and imaging conditions, achieving 98.46% accuracy and a PR-AUC of 0.9937 in ALL classification.

Existing instance segmentation evaluation metrics—such as AP@50, PQ, SEG, and AJI—are limited in biological imaging due to their reliance on hard IoU thresholds, object-level normalization biases, and greedy matching strategies, which lead to discontinuous scores, low sensitivity, and unreliable ranking. This work proposes Maximum Matching Accuracy (MMA), a threshold-free and continuous metric that establishes a globally optimal one-to-one correspondence between predicted and ground-truth instances via the Hungarian algorithm and quantifies total overlap using pixel-level normalization. Experimental results demonstrate that MMA substantially outperforms existing metrics across synthetic failure cases, progressive perturbations, and model ranking tasks, exhibiting superior stability, sensitivity, and interpretability.

This work addresses the significant performance degradation of conventional cell instance segmentation models on out-of-distribution cell types and the inefficiency of existing interactive methods that require per-instance prompting. The authors propose a novel Group Prompting paradigm, enabling interaction at the cell-type level rather than the instance level: a single click per cell type suffices to segment all instances of that type. Built upon a Chain-of-Prompts framework, the method recursively expands reliable prompt points on the frozen multi-scale image features of SAM, leveraging their inherent ability to aggregate similar cells. This reduces interaction complexity from O(N) to O(T), where N is the number of instances and T the number of types. Experiments demonstrate that with just one click per type, the approach retains over 90% of per-instance performance on three annotated datasets and exceeds 99% on four homogeneous datasets—surpassing fully supervised baselines without any additional training.

Manual annotation of cells and structures in whole-slide tissue images is prohibitively expensive. This work proposes an end-to-end “cluster-then-annotate” pipeline that dramatically reduces labeling effort by shifting the annotation target from individual structures to representative clusters. The method leverages Cellpose-SAM for structural segmentation, ResNet-50 to extract morphological embeddings, and UMAP followed by DBSCAN for dimensionality reduction and clustering. Annotators label only one exemplar per cluster, enabling comprehensive coverage of the entire image. Evaluated across 13 tissue types and 3,696 annotated structures, the approach achieves a weighted cluster-to-label alignment accuracy of 96.8%, with perfect consistency in seven tissue categories. This framework demonstrates the first successful generalization of morphology-based clustering across species and diverse tissue types, substantially enhancing annotation efficiency.