CSRefiner: a lightweight framework for fine-tuning cell segmentation models with small datasets

Simple SummaryThis study evaluates the performance degradation of machine learning models for segmenting gliomas in brain magnetic resonance images caused by domain shift and proposed possible solutions. Although machine learning models exhibit significant potential for clinical applications, performance degradation in different cohorts is a problem that must be solved. In this study, we identify the impact of the performance degradation of machine learning models to be significant enough to render clinical applications difficult. This demonstrates that it can be improved by fine-tuning methods with a small number of cases from each facility, although the data obtained appeared to be biased. Our method creates a facility-specific machine learning model from a small real-world dataset and public dataset; therefore, our fine-tuning method could be a practical solution in situations where only a small dataset is available.Machine learning models for automated magnetic resonance image segmentation may be useful in aiding glioma detection. However, the image differences among facilities cause performance degradation and impede detection. This study proposes a method to solve this issue. We used the data from the Multimodal Brain Tumor Image Segmentation Benchmark (BraTS) and the Japanese cohort (JC) datasets. Three models for tumor segmentation are developed. In our methodology, the BraTS and JC models are trained on the BraTS and JC datasets, respectively, whereas the fine-tuning models are developed from the BraTS model and fine-tuned using the JC dataset. Our results show that the Dice coefficient score of the JC model for the test portion of the JC dataset was 0.779 ± 0.137, whereas that of the BraTS model was lower (0.717 ± 0.207). The mean Dice coefficient score of the fine-tuning model was 0.769 ± 0.138. There was a significant difference between the BraTS and JC models (p < 0.0001) and the BraTS and fine-tuning models (p = 0.002); however, no significant difference between the JC and fine-tuning models (p = 0.673). As our fine-tuning method requires fewer than 20 cases, this method is useful even in a facility where the number of glioma cases is small.

Advances in regenerative medicine highlighted the need for label-free cell image analysis to replace conventional microscopic observation for non-invasive cell quality evaluation. Image-based evaluation provides an efficient, quantitative, and automated approach to cell analysis, but segmentation remains a critical and challenging step. In this study, we investigated how training dataset design influenced the robustness of U-Net models for cell segmentation, focusing on challenges posed by limited data availability in cell culture. Using 2592 image pairs from four cell types representing key morphological categories, we constructed 42 investigation patterns to evaluate the effects of dataset size, dataset content, and morphological diversity on model performance. Our results showed that robust segmentation models could be developed with approximately 10 raw images captured using a 4× objective lens, a much smaller dataset than typically assumed. The dataset content was found to be crucial: training dataset images that captured commonly observed cell patterns yielded more robust models compared to those capturing rare or irregular cell patterns, which often impaired model performance with large deviations. Additionally, including both spindle and round cell morphologies in the training datasets improved model robustness when tested across all four cell types, while datasets restricted to a single morphology type could not achieve robust models. These findings highlight the importance of curating datasets that capture representative yet diverse cell morphologies. By addressing critical questions about dataset design, this study provides actionable guidance for the effective use of deep learning-based cell segmentation models in manufacturing and research applications.

Mammalian reproduction requires the fusion of two specialized cells: an oocyte and a sperm. In addition to producing gametes, the reproductive system also provides the environment for the appropriate development of the embryo. Deciphering the reproductive system requires understanding the functions of each cell type and cell-cell interactions. Recent single-cell omics technologies have provided insights into the gene regulatory network in discrete cellular populations of both the male and female reproductive systems. However, these approaches cannot examine how the cellular states of the gametes or embryos are regulated through their interactions with neighboring somatic cells in the native tissue environment owing to tissue disassociations. Emerging spatial omics technologies address this challenge by preserving the spatial context of the cells to be profiled. These technologies hold the potential to revolutionize our understanding of mammalian reproduction. We aim to review the state-of-the-art spatial transcriptomics (ST) technologies with a focus on highlighting the novel biological insights that they have helped to reveal about the mammalian reproductive systems in the context of gametogenesis, embryogenesis, and reproductive pathologies. We also aim to discuss the current challenges of applying ST technologies in reproductive research and provide a sneak peek at what the field of spatial omics can offer for the reproduction community in the years to come. The PubMed database was used in the search for peer-reviewed research articles and reviews using combinations of the following terms: 'spatial omics', 'fertility', 'reproduction', 'gametogenesis', 'embryogenesis', 'reproductive cancer', 'spatial transcriptomics', 'spermatogenesis', 'ovary', 'uterus', 'cervix', 'testis', and other keywords related to the subject area. All relevant publications until April 2023 were critically evaluated and discussed. First, an overview of the ST technologies that have been applied to studying the reproductive systems was provided. The basic design principles and the advantages and limitations of these technologies were discussed and tabulated to serve as a guide for researchers to choose the best-suited technologies for their own research. Second, novel biological insights into mammalian reproduction, especially human reproduction revealed by ST analyses, were comprehensively reviewed. Three major themes were discussed. The first theme focuses on genes with non-random spatial expression patterns with specialized functions in multiple reproductive systems; The second theme centers around functionally interacting cell types which are often found to be spatially clustered in the reproductive tissues; and the thrid theme discusses pathological states in reproductive systems which are often associated with unique cellular microenvironments. Finally, current experimental and computational challenges of applying ST technologies to studying mammalian reproduction were highlighted, and potential solutions to tackle these challenges were provided. Future directions in the development of spatial omics technologies and how they will benefit the field of human reproduction were discussed, including the capture of cellular and tissue dynamics, multi-modal molecular profiling, and spatial characterization of gene perturbations. Like single-cell technologies, spatial omics technologies hold tremendous potential for providing significant and novel insights into mammalian reproduction. Our review summarizes these novel biological insights that ST technologies have provided while shedding light on what is yet to come. Our review provides reproductive biologists and clinicians with a much-needed update on the state of art of ST technologies. It may also facilitate the adoption of cutting-edge spatial technologies in both basic and clinical reproductive research.

Deep learning-based segmentation models can accelerate the analysis of high-throughput microscopy data by automatically identifying and classifying cells in images. However, the datasets needed to train these models are typically assembled via laborious hand-annotation. This limits their scale and diversity, which in turn limits model performance. We present Cell-APP (Cellular Annotation and Perception Pipeline), a tool that automates the annotation of high-quality training data for transmitted-light (TL) cell segmentation. Cell-APP uses two inputs-paired TL and nuclear fluorescence images-and operates in two main steps. First, it extracts each cell's location from the nuclear fluorescence channel and provides these locations to promptable deep learning models to generate cell masks. Then, it classifies each cell as mitotic or nonmitotic based on nuclear features. Together, these masks and classifications form the basis for cell segmentation training data. By training vision-transformer-based models on Cell-APP-generated datasets, we demonstrate how Cell-APP enables the creation of both cell line-specific and multi-cell line segmentation models. Cell-APP thus empowers laboratories to tailor cell segmentation models to their needs and outlines a scalable path to creating general models for the research community.

Deep learning-based segmentation models can accelerate the analysis of high-throughput microscopy data by automatically identifying and classifying cells in images. However, the datasets needed to train these models are typically assembled via laborious hand-annotation. This limits their scale and diversity, which in turn limits model performance. We present Cell-APP (Cellular Annotation and Perception Pipeline), a tool that automates the annotation of high-quality training data for transmitted-light (TL) cell segmentation. Cell-APP uses two inputs—paired TL and nuclear fluorescence images—and operates in two main steps. First, it extracts each cell’s location from the nuclear fluorescence channel and provides these locations to promptable deep learning models to generate cell masks. Then, it classifies each cell as mitotic or non-mitotic based on nuclear features. Together, these masks and classifications form the basis for cell segmentation training data. By training vision-transformer-based models on Cell-APP-generated datasets, we demonstrate how Cell-APP enables the creation of both cell line-specific and multi-cell line segmentation models. Cell-APP thus empowers laboratories to tailor cell segmentation models to their needs, and outlines a scalable path to creating general models for the research community.

A 9.4T brain MRI is the highest resolution MRI scanner in the public market. It offers submillimeter brain imaging with exceptional anatomical detail, making it one of the most powerful tools for detecting subtle structural changes associated with neurological conditions. Current segmentation models are optimized for lower-field MRI (1.5T–3T), and they struggle to perform well on 9.4T data. In this study, we present the GA-MS-UNet++, the world’s first deep learning-based model specifically designed for 9.4T brain MRI segmentation. Our model integrates multi-scale residual blocks, gated skip connections, and spatial channel attention mechanisms to improve both local and global feature extraction. The model was trained and evaluated on 12 patients in the UltraCortex 9.4T dataset and benchmarked against four leading segmentation models (Attention U-Net, Nested U-Net, VDSR, and R2UNet). The GA-MS-UNet++ achieved a state-of-the-art performance across both evaluation sets. When tested against manual, radiologist-reviewed ground truth masks, the model achieved a Dice score of 0.93. On a separate test set using SynthSeg-generated masks as the ground truth, the Dice score was 0.89. Across both evaluations, the model achieved an overall accuracy of 97.29%, precision of 90.02%, and recall of 94.00%. Statistical validation using the Wilcoxon signed-rank test (p < 1 × 10−5) and Kruskal–Wallis test (H = 26,281.98, p < 1 × 10−5) confirmed the significance of these results. Qualitative comparisons also showed a near-exact alignment with ground truth masks, particularly in areas such as the ventricles and gray–white matter interfaces. Volumetric validation further demonstrated a high correlation (R2 = 0.90) between the predicted and ground truth brain volumes. Despite the limited annotated data, the GA-MS-UNet++ maintained a strong performance and has the potential for clinical use. This algorithm represents the first publicly available segmentation model for 9.4T imaging, providing a powerful tool for high-resolution brain segmentation and driving progress in automated neuroimaging analysis.

Conventional scRNA-seq expression analyses rely on the availability of a high quality genome annotation. Yet, as we show here with scRNA-seq experiments and analyses spanning human, mouse, chicken, mole rat, lemur and sea urchin, genome annotations are often incomplete, in particular for organisms that are not routinely studied. To overcome this hurdle, we created a scRNA-seq analysis routine that recovers biologically relevant transcriptional activity beyond the scope of the best available genome annotation by performing scRNA-seq analysis on any region in the genome for which transcriptional products are detected. Our tool generates a single-cell expression matrix for all transcriptionally active regions (TARs), performs single-cell TAR expression analysis to identify biologically significant TARs, and then annotates TARs using gene homology analysis. This procedure uses single-cell expression analyses as a filter to direct annotation efforts to biologically significant transcripts and thereby uncovers biology to which scRNA-seq would otherwise be in the dark.

The automatic segmentation of medical images has widespread applications in modern clinical workflows. The Segment Anything Model (SAM), a recent development of foundational models in computer vision, has become a universal tool for image segmentation without the need for specific domain training. However, SAM's reliance on prompts necessitates human-computer interaction during the inference process. Its performance on specific domains can also be limited without additional adaptation. In contrast, traditional models like nnUNet are designed to perform segmentation tasks automatically during inference and can work well for each specific domain, but they require extensive training on domain-specificdatasets. To leverage the advantages of both foundational and domain-specific models and achieve fully automated segmentation with limited training samples, we propose nnSAM, which combines the robust feature extraction capabilities of SAM with the automatic configuration abilities of nnUNet to enhance the accuracy and robustness of medical image segmentation on smalldatasets. We propose the nnSAM model for small sample medical image segmentation. We made optimizations for this goal via two main approaches: first, we integrated the feature extraction capabilities of SAM with the automatic configuration advantages of nnUNet, which enables robust feature extraction and domain-specific adaptation on small datasets. Second, during the training process, we designed a boundary shape supervision loss based on level set functions and curvature calculations, enabling the model to learn anatomical shape priors from limited annotationdata. We conducted quantitative and qualitative assessments on the performance of our proposed method on four segmentation tasks: brain white matter, liver, lung, and heart segmentation. Our method achieved the best performance across all tasks. Specifically, in brain white matter segmentation using 20 training samples, nnSAM achieved the highest DICE score of 82.77 ( 10.12) % and the lowest average surface distance (ASD) of 1.14 ( 1.03) mm, compared to nnUNet, which had a DICE score of 79.25 ( 17.24) % and an ASD of 1.36 ( 1.63) mm. A sample size study shows that the advantage of nnSAM becomes more prominent under fewer trainingsamples. A comprehensive evaluation of multiple small-sample segmentation tasks demonstrates significant improvements in segmentation performance by nnSAM, highlighting the vast potential of small-samplelearning.

With the aim to advance the understanding of immune regulation in MCL and to identify targetable T-cell subsets, we set out to combine image analysis and spatial omic technology focused on both early and late differentiation stages of T cells. MCL patient tissue (n = 102) was explored using image analysis and GeoMx spatial omics profiling of 69 proteins and 1812 mRNAs. Tumor cells, T helper (TH) cells and cytotoxic (TC) cells of early (CD57-) and late (CD57+) differentiation stage were analyzed. An image analysis workflow was developed based on fine-tuned Cellpose models for cell segmentation and classification. TC and CD57+ subsets of T cells were enriched in tumor-rich compared to tumor-sparse regions. Tumor-sparse regions had a higher expression of several key immune suppressive proteins, tentatively controlling T-cell expansion in regions close to the tumor. We revealed that T cells in late differentiation stages (CD57+) are enriched among MCL infiltrating T cells and are predictive of an increased expression of immune suppressive markers. CD47, IDO1 and CTLA-4 were identified as potential targets for patients with T-cell-rich MCL TIME, while GITR might be a feasible target for MCL patients with sparse T-cell infiltration. In subgroups of patients with a high degree of CD57+ TC-cell infiltration, several immune checkpoint inhibitors, including TIGIT, PD-L1 and LAG3 were increased, emphasizing the immune-suppressive features of this highly differentiated T-cell subset not previously described in MCL.

Focused ion beam scanning electron microscopy (FIB-SEM) tomography has increasingly been utilized for acquiring three-dimensional (3D) microstructure features at the sub-micron scale in irradiated nuclear materials. This technique involves sequential ion beam slicing followed by electron beam imaging and compositional mapping using energy dispersive spectroscopy (EDS). Despite its growing use, several challenges persist. These include the time-intensive nature of data collection of EDS data, difficulties in distinguishing between various microstructures, and issues with image alignment. These challenges currently limit the broader application of FIB-SEM tomography in the field. To overcome these limitations, we propose using convolutional neural networks (CNNs) to automate microstructure identification in SEM images. Our study introduces a new framework for identifying microstructures in irradiated U-10Zr (wt%) metallic fuel with limited annotated data. The framework includes the creation of a reliable annotated dataset with paired SEM and ground truth data from EDS maps, the applications of CNNs for microstructure identification, and the validation of model performance. Specifically, we employed the Segment Anything Model (SAM) to align SEM images with corresponding EDS maps and focused ion beam (FIB) tomography SEM data. We evaluate several models, including Patch-based U-Net, Attention U-Net, and Residual U-Net, finding that patch-based U-Net exhibits superior segmentation performance and consistency. This approach reduces reliance on EDS detectors and aids in accelerating nuclear material analysis process, highlighting the potential of advanced deep learning techniques to improve microstructural understanding in nuclear material. This is the first framework to integrate SAM and Patch-based CNN models for semantic segmentation of irradiated nuclear materials, with potential applicability to other tomography datasets.

Practical guidelines for cell segmentation models under optical aberrations in microscopy

BackgroundDeciphering the tumor-immune microenvironment (TIME) and the impact of cell-to-cell interaction for immunotherapy response is pivotal for developing biomarker signatures for patient stratification and novel therapies. Spatial phenotyping and functional...

Spatial omics technologies generate high-dimensional, spatially resolved molecular data across transcripts, proteins, metabolites and lipids, requiring computational models that account for tissue topology, multi-scale organization, and experimental noise. Although machine-learning (ML) and deep-learning (DL) methods have rapidly proliferated to meet these demands, the field still lacks clear methodological guidance for selecting models adapted to specific spatial constraints and biological questions. Here, we provide a critical and comparative synthesis of ML/DL approaches across core spatial omics tasks, including batch-effect correction, resolution enhancement, tissue and cell segmentation, spatial domain discovery, cell-type deconvolution, and model interpretability. Classical ML methods such as clustering, random forests, and other ensemble classifiers, offer interpretable baselines but are limited in their capacity to model non-linear spatial dependencies. Modern DL architectures, including convolutional and graph neural networks, transformers and generative models, capture complex spatial patterns and support multi-omics integration, yet face persistent challenges related to data scarcity, annotation burden, computational cost, and uncertainty estimation. Emerging strategies such as optimal transport, cross-modal attention, graph-linked embeddings, and foundation models enhance cross-modality alignment but require rigorous evaluation of their assumptions and operational constraints. We further discuss practical solutions, including self-supervised pretraining, federated learning and the adoption of standardized spatial data formats, to enhance scalability, reproducibility, and clinical readiness. Finally, we propose a decision framework that highlights when specific ML/DL families are most suitable according to data modality, spatial resolution, tissue architecture, and intended clinical application. By integrating methodological critique with actionable recommendations, this review offers a roadmap for the reproducible, interpretable, and clinically translatable deployment of ML and DL models in spatial omics.

White matter hyperintensities (WMH) represent neuroimaging markers of cerebral small vessel disease. We aimed to develop and validate a deep learning-based simultaneous, automatic WMH segmentation and classification model in patients with cognitive impairment. This retrospective study included images from consecutive patients with cognitive impairment from a tertiary hospital. A segmentation model was trained and tuned on 448 and 149 subjects. A classification compartment inherited from the segmentation model, utilizing the multi-task multiple instance learning framework (MTMIL), was trained and tuned on 1186 and 394 subjects. Tests of segmentation and classification tasks were performed on 149 and 394 subjects in the internal testing dataset and 100 subjects in the external dataset. We evaluated five different models to select the segmentation branch.Classification according to the Fazekas scale used three categories (normal/mild, moderate, severe), and locations were separately reported as periventricular and deep WMH. For classification evaluation, three experienced neuroradiologists evaluated test datasets. Between January 2016 and December 2019, 1974 consecutive patients (mean age 71.1 ± 9.7) were included. Dice score performance of the UNet with Resnet-34 encoder model on internal and external testing datasets was 0.88 (95% CI: 0.88-0.89) and 0.85 (95% CI: 0.84-0.86) for single-class segmentation, and 0.77 (95% CI: 0.76-0.79) and 0.72 (95% CI: 0.71-0.74) for multi-class segmentation. The accuracy of the Fazekas scale classification model was 0.88 and 0.87 at periventricular and deep WMH with internal datasets and 0.68 and 0.75 with external datasets. These results demonstrate the high segmentation and classification performance of our models and their potential for deployment as accurate diagnostic support tools for quantified evaluation of WMH. Question We developed a deep learning-based, simultaneous, automatic white matter hyperintensity (WMH) segmentation and classification model using data from patients with cognitive impairment. Findings Segmentation model achieved Dice scores of 0.72-0.88 for single-class and multi-class segmentation. Fazekas score classification accuracy ranged from 0.68 to 0.88 for periventricular/deep WMH. Clinical relevance Our study demonstrated high segmentation and classification performance of deep learning-based models and potential for deployment as accurate diagnostic support tools for quantified evaluation of white matter hyperintensities.

Introduction Wildfire detection and segmentation play a critical role in environmental monitoring and disaster prevention. However, existing deep learning-based segmentation models often struggle to identify wildfire boundaries accurately due to complex image features and limited annotated data. Methods We propose a novel segmentation network called PPCNet, which integrates three key modules: a Panoramic Feature Fusion (PFF) module for multi-scale feature extraction, a Dense Feature Fusion Encoder (DFFE) to capture contextual details, and a Local Detail Compensation (LDC) loss function to enhance boundary accuracy. Additionally, we design a pseudo-label optimization framework to leverage unlabeled data effectively. Results Experiments were conducted on multiple wildfire datasets, and the results show that PPCNet achieves superior performance compared to state-of-the-art methods. Our model demonstrates significant improvements in segmentation accuracy and boundary localization, validated through quantitative metrics and visual comparisons. Discussion The integration of PFF, DFFE, and LDC components enables PPCNet to generalize well across different wildfire scenarios. The use of pseudo-labeling further enhances performance without requiring additional labeled data, making it suitable for real-world deployment in wildfire monitoring systems.