SparseXMIL: Leveraging sparse convolutions for context-aware and memory-efficient classification of whole slide images in digital pathology.

Use of whole slide images for primary pathology evaluation of non-clinical toxicology studies in a GLP-validated environment.

PTCMIL: multiple instance learning via prompt token clustering for whole slide image analysis.

Adaptive deep learning for slide-level multi-label biomarker prediction in breast cancer whole slide images via misprediction risk analysis

A content-aware variable-rate framework for pathology learned image compression (PathoLIC).

The integration of artificial intelligence into pathology is transforming the assessment of histological and immunohistochemical (IHC) slides, offering opportunities to reduce variability and streamline diagnostics. In practical terms, most available tools and research models emulate the diagnostic capabilities of pathologists by detecting, grading, and classifying tumours and other diseases. More recent applications have moved beyond mimicry, aiming to predict established biomarkers, such as microsatellite instability or IHC-based markers, and to tackle even more ambitious tasks, such as directly predicting patient prognosis from H&E whole slide images. Remarkably, novel computational tools are now being designed as companion diagnostic assays, linking the automated evaluation of specific IHC biomarkers to the prediction of response to specific drugs, potentially marking a new chapter in the evolution of digital and computational pathology. The TROPION-PanTumor01 trial recently demonstrated the superiority of a supervised machine learning model (termed the quantitative continuous score [QCS] by the vendor) in assessing TROP2 IHC compared with human scoring, promising better stratification of patients with non-small cell lung cancer for treatment with datopotamab deruxtecan. The same approach has shown promise in refining HER2 (human epidermal growth factor receptor 2) and PD-L1 (programmed death-ligand 1) evaluations, revealing patient subgroups that may benefit from targeted therapies. Moreover, other similar approaches are progressively reaching the market, posing significant opportunities and challenges for clinicians involved in the care of patients with cancer. This Perspective is promoted by the European Society of Digital and Integrative Pathology (ESDIP, founded in 2016, and having long-standing experience in computational pathology, esdipath.org) and the European Interdisciplinary Society of Artificial Intelligence for Cancer Research (ESAC, a recently established initiative, founded in 2024, esac-network.eu), both bringing together clinicians, engineers and other professionals dedicated to the development and clinical translation of computational approaches aimed at improving patient care. It aims to provide an informed overview of novel computational pathology companion diagnostic tools, with a particular focus on the background that practicing pathologists and oncologists need to have with these tools, when transitioning from research to clinical practice, irrespective of their prior familiarity with computational approaches. © 2026 The Author(s). The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

In computational pathology, the gigapixel scale of Whole-Slide Images (WSIs) requires their decomposition into thousands of patches, resulting in high-dimensional embeddings that are computationally costly to process and often dominated by uninformative regions. Existing patch selection methods typically rely on heuristic sampling and do not explicitly address the trade-off between representation compactness and diagnostic accuracy. To address this gap, we propose EvoPS (Evolutionary Patch Selection), a novel framework that formulates patch selection within the training embedding space as a multi-objective optimization problem and leverages an evolutionary search to simultaneously minimize the number of selected patch embeddings and maximize the performance of a downstream similarity search task, generating a Pareto front of optimal trade-off solutions. By identifying a compact and diagnostically informative subset of training patches, EvoPS produces higher-quality training representations that reduce memory requirements and improve the signal-to-noise ratio of the training set. We validated our framework across four major cancer cohorts from The Cancer Genome Atlas (TCGA) using five histopathology foundation models. The results demonstrate that EvoPS can reduce the required number of training patches by over 90% while consistently maintaining or even improving the final classification F1-score compared to a state-of-the-art patch selection method. The EvoPS framework provides a robust and principled method for creating efficient, accurate, and interpretable WSI representations, empowering users to select an optimal balance between computational cost and diagnostic performance.

MorphoNet: Morphological sub-region-based structure learning for WSI analysis.

Towards generalizable pathology reports via a multimodal LLM with the multicenter in-context learning.

It is imperative to identify patients with prostate cancer (PCa) who will not benefit from androgen receptor signaling inhibitors and to improve their clinical outcomes. Using artificial intelligence (AI), in this multicenter cohort study of 623 PCa patients, we identified 13 cellular morphometric biomarkers (CMBs), as a New Approach Methodology (NAM), from whole slide images of needle biopsies in clinical trial specimens (NCT02430480, n=37) that accurately predicted response to neoadjuvant androgen deprivation therapy (NADT) plus enzalutamide (AUC: 0.981, 95% CI [0.979, 0.983]). Importantly, the 13-CMB model stratified PCa patients into responders and non-responders after NADT across two independent hospital cohorts. In one cohort (n=122), the model identified groups with significantly different pathologic complete response (pCR) (p=0.0005) and biochemical recurrence-free survival (BCRFS) (p=0.024). In the second cohort (n=60), the model similarly distinguished patients with significantly different BCRFS (p=0.031). The 13-CMB model also stratified PCa patients in the TCGA-PRAD cohort (n=396) with distinct progression-free survival (p=0.0017). Importantly, across hospital cohorts and the TCGA-PRAD cohort, the 13-CMB model demonstrated significant and independent clinical value after adjustment for established clinical factors and commonly used genomic biomarkers, including Decipher and Oncotype DX. Furthermore, CMBs accurately predicted the molecular differences between stratified patient groups and the potential benefit from mTOR inhibitors in non-responders, which were validated through IHC staining and patient-derived organoids (n=8), respectively. Overall, our AI-powered CMB model, relying only on routine needle biopsy specimens, could potentially serve as a robust solution for precision management of PCa patients.

The classification of immunophenotypes in muscle-invasive bladder cancer (MIBC) is critical for predicting immunotherapy response and clinical outcomes, yet current assessment methods lack standardization and scalability. We developed and validated an artificial intelligence-based MIBC Immunophenotype Diagnostic System using computational pathology to enable reproducible classification from routine hematoxylin and eosin-stained whole-slide images. In this multicenter retrospective diagnostic study, consecutive patients who underwent partial or radical cystectomy between 2014 and 2024 from two Chinese hospitals and The Cancer Genome Atlas cohort were included, with an independent cohort receiving immune checkpoint inhibitors for treatment efficacy evaluation. The system integrates Hover-Net-based nuclear classification with cell structure graph networks to model spatial cellular interactions within the tumor microenvironment. Across external validation cohorts, the model achieved macro-area under the curve values of 0.922-0.956 and macro-accuracy of 0.922-0.950, demonstrating robust generalizability. In a human-AI collaboration study, the system outperformed junior and senior pathologists and significantly improved junior pathologists' diagnostic accuracy while reducing review time. Predicted Inflamed tumors exhibited enriched CD8+ T-cell infiltration, elevated checkpoint gene expression, and stronger correlation with immunotherapy response. These findings support clinical translation for precision immuno-oncology in bladder cancer.

A cell comparative multiple instance learning network guided by image quality assessment for cervical whole slide image classification.

Many liver diseases have a distinct zonation pattern. Similarly, most metabolic processes in the hepatic lobule are also spatially organized. Understanding the interplay between a zonated disease pattern and its impact on zonated metabolic liver function requires the joint quantification of both phenomena. Our study presents an image analysis workflow for the joint zonated quantification of multiple parameters from whole-slide images of conventionally stained serial sections of mouse livers. As a proof of concept, we used small stacks of six adjacent sections, differentially stained with HE, GS and four different CYP enzymes from three mice with different severities of steatosis. Portal fields and central veins were annotated and transferred to adjacent slide images via image registration. The result was visually confirmed to avoid errors in the non-rigid transformation process. This approach allowed identifying the same lobules and zones in multiple consecutive sections. Zones were obtained by arbitrarily dividing the distance between portal fields and central veins into twelve intralobular zones. Zonal distributions of various parameter combinations quantified from multiple slides were visualized in the geometry of lobules and zones and plotted as scatter diagrams. Using this workflow, we could visualize differences in the heterogeneous expression patterns of marker proteins in normal and steatotic livers. As a next step, the algorithm presented here can be applied to a scientific question such as quantifying the impact of zonated steatosis of different severity on the zonated expression pattern and the resulting metabolic function of CYP enzymes.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-46721-5.

Tissue detection is a crucial first step in most digital pathology applications. By applying image segmentation algorithms, all tissue is delineated and background discarded from further analyses, improving both computational efficiency and analytical results. Details of the segmentation algorithm are rarely reported, and there is a lack of studies investigating the downstream effects of a poor segmentation algorithm. Disregarding tissue detection quality could jeopardize patient safety if diagnostically relevant parts of the specimen are excluded from analysis in clinical applications. This study aims to determine whether performance of downstream tasks is sensitive to the tissue detection method, and to compare the performance of a classical and an AI-based tissue detection approach. To this end, we trained an AI model for Gleason grading of prostate cancer in whole slide images (WSIs) using two different tissue detection algorithms: thresholding (classical) and UNet++ (AI). A total of 33,823 WSIs scanned on seven digital pathology scanners were used to train the segmentation AI model. The downstream Gleason grading algorithm was trained and tested using 70,524 WSIs from 13 clinical sites scanned on 13 different scanners. On the slides where tissue could be detected by both algorithms, no significant difference in overall Gleason grading performance was observed. There was a decrease from 118 (0.43%) to 24 (0.09%) fully undetected tissue samples when switching from thresholding-based tissue detection to AI-based, suggesting this AI model may be more reliable than the classical model for avoiding total failures on slides with unusual appearance. Moreover, tissue detection dependent clinically significant variations in AI grading were observed in 3.5% of malignant slides, highlighting the role of tissue detection for optimal clinical performance of diagnostic AI.

Accurate prediction of Homologous Recombination Deficiency (HRD) is vital for personalized cancer therapy, yet genomic assays are often costly and complex. Predicting HRD from Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs) via deep learning offers an alternative, but model generalization across diverse cancer types remains challenging. We developed a three-stage transfer learning framework to improve HRD prediction across cancer types. After establishing baseline models on data from eight cancer types, the TCGA-BRCA model was selected as a pretrained model for transfer learning applied to eligible cohorts. Transfer learning enhanced HRD prediction performance specifically in cancers with histological similarity to the TCGA-BRCA source model. The AUC increased by + 4% (TCGA-LUAD), + 5% (CPTAC-LUAD), and + 1% (SARC) compared to baseline models. Interpretability analysis confirmed that the model's predictions were driven by histologically relevant tumor regions, which was further supported by quantitative cellular analysis showing that HRD-high status was characterized by a significantly higher density of neoplastic cells (HRD-high: median 16.0, IQR 1.0-27.0 vs. HRD-low: median 1.0, IQR:0.0-5.0; p < 0.001), whereas HRD-low cases featured a stromal-rich microenvironment. Conclusions This study demonstrates that transfer learning can enhance HRD prediction from H&E images for cancer types sharing histopathological features with the TCGA-BRCA source model, offering a more efficient and accessible approach for clinical HRD assessment.

Accurate prediction of recurrence risk is essential to devise effective and personalized treatment strategies for patients with soft tissue sarcoma (STS). This study aimed to develop and validate a multimodal deep learning framework that integrates clinical features, preoperative MR images, and hematoxylin and eosin-stained whole slide images (WSIs) to predict recurrence in patients with STS. A total of 323 patients with STS were retrospectively enrolled from two hospitals, serving as development and validation sets, respectively. The ShuffleNetV2 network was utilized to develop patch-level and WSI-level signatures. A convolutional neural network fusing the channel and spatial attention mechanisms was used to develop a radiology signature. The combined model was built by integrating clinical features, radiology signature score, and WSI-level signature score with Cox regression analysis. The combined model demonstrated superior performance in the validation set, achieving a C-index of 0.857 and a time-dependent area under the curve of 0.959. Class activation maps facilitated the monitoring of suspected regions to inform recurrence decisions. The recurrence-free survival times of the low- and high-risk cohorts were statistically different (p < 0.05). The proposed multimodal framework offers satisfactory accuracy for predicting recurrence risk in patients with STS and could guide the choice of treatment modality.

Artificial intelligence (AI) has demonstrated high accuracy in detecting lymph node (LN) metastases in treatment-naïve invasive breast cancer. However, its performance post-neoadjuvant chemotherapy (NACT) remains underexplored, where therapy-related morphologic alterations complicate assessment. To evaluate an AI algorithm (aetherAI) for detecting LN metastases in post-NACT invasive breast cancer. This retrospective study included 72 patients (58 [80.5%] post-NACT; 14 [19.5%] treatment-naïve), yielding 526 whole slide images and 1290 LNs. AI-generated LN contours were reviewed and modified, if needed. LN status was independently assessed by 4 pathologists and the AI algorithm. Discordant cases were resolved by consensus/immunohistochemistry to define ground truth, against which AI was compared. AI accurately contoured 1089 of 1290 LNs (84.4%). Pathologists showed 94.42% agreement, with 5.58% discordance. Against consensus ground truth, AI achieved 94.06% sensitivity, 90.14% specificity, 91.47% accuracy, and area under the curve (AUC) of 0.92. Performance was superior in treatment-naïve LNs (sensitivity, 98.95%; specificity, 98.08%; AUC, 0.99) compared with post-NACT LNs (sensitivity, 92.71%; specificity, 88.36%; AUC, 0.91) (P < .001). Within the post-NACT group, significantly reduced accuracy was associated with interobserver discordance (P < .001) and smaller metastases (micrometastasis, isolated tumor cells) (P < .001), but not noted in treatment-naïve cases. Most false results were attributable to therapy-induced changes. Unlike most prior studies restricted to treatment-naïve LNs, this post-NACT cohort demonstrates that NACT-related alterations reduce AI sensitivity and accuracy. Our findings suggest that incorporating post-NACT nodes into algorithm training will be essential for reliable clinical translation of AI-assisted LN evaluation.

Distant metastasis (DM) is the primary driver of cancer-related mortality, and its clinical prediction remains challenging due to the lack of robust biomarkers. This study proposes a novel graph representation that effectively identifies discriminative morphological features from histopathological whole slide images (WSIs). By transforming high-resolution WSIs into topological graphs, the proposed method leverages graph neural networks (GNNs) to capture complex spatial dependencies and cellular organizations critical for metastatic progression. The study is evaluated on a large-scale pan-cancer dataset and demonstrates superior performance in distilling shared metastatic patterns across diverse malignancies. Furthermore, the cross-dataset robustness of this representation is validated by training on a specialized nasopharyngeal carcinoma cohort (TJ-NPC) and evaluating on independent public datasets. The results highlight the potential of computational pathology to provide scalable, objective risk stratification, offering a high accuracy tool for personalized clinical intervention.

The OncotypeDX 21-gene assay guides adjuvant chemotherapy decisions in early-stage, hormone receptor-positive, HER2-negative breast cancer, but cost and turnaround time limit access. This study presents a deep learning-based approach for predicting OncotypeDX recurrence scores directly from hematoxylin and eosin-stained whole slide images. Our approach leverages a deep learning foundation model pre-trained on 171,189 slides via self-supervised learning, which is fine-tuned for our task. The model was developed and validated using five independent cohorts, out of which three are external. On the two external cohorts that include OncotypeDX scores, the model achieved an AUC of 0.836 and 0.817, and identified 22% and 16.3% of the patients as low-risk with sensitivity of 0.97 and 0.97 and negative predictive value of 0.97 and 0.96, showing strong generalizability despite variations in staining protocols and imaging devices. Kaplan-Meier analysis demonstrated that patients classified as low-risk by the model had a significantly better prognosis than those classified as high-risk, with a hazard ratio of 4.1 (P < 0.001) and 2.0 (P < 0.01) on the two external cohorts that include patient outcomes. This artificial intelligence-driven solution offers a rapid, cost-effective, and scalable alternative to genomic testing, with the potential to enhance personalized treatment planning, especially in resource-constrained settings.

Digital pathology involves the digitisation of histology slides, which is vital in modern healthcare, especially for cancer detection and diagnosis. Still, the manual analysis of these digital slides by expert pathologists is time-consuming and prone to inconsistencies. This workload often results in subjective grading and a lack of agreement between different pathologists (inter-operator variability) and even with the same pathologist over time (intra-operator variability). With the advent of whole slide imaging (WSI) and high computational ability, the application of deep learning (DL), particularly convolutional neural networks, has developed rapidly in the past decades in the domain of digital pathology. Therefore, this study introduces a Hybrid Temporal Deep Learning Network for Automated Malignant Cell Classification in Cytology Slides (HTDL-MCCCS) model. The objective is to improve diagnostic consistency, reduce workload, and enable scalable screening in resource-limited clinical settings. The HTDL-MCCCS model consists of four integrated stages: First, slide pre-processing is performed using stain normalisation, artefact removal, patch extraction, and nuclei segmentation to enhance morphological clarity. Second, the pyramid vision transformer (PVT) is utilised to learn robust cell-level features from unlabeled slide patches. Third, the bidirectional temporal recurrent unit (BiTRU) is employed for the automated classification of malignant and benign cells. Lastly, a detailed analysis using Grad-CAM is included to support clinical interpretability and trust. A wide-ranging simulation analysis of the HTDL-MCCCS methodology portrayed superior outcomes of 98.16% and 96.42% under the SIPaKMeD and Herlev datasets.