Accurate molecular classification of breast core needle biopsy (CNB) tissue is important for determining neoadjuvant systemic therapies for invasive breast cancer. The researchers aimed to evaluate the concordance rate (CR) of molecular subtypes between CNBs and surgical specimens. This study was conducted with invasive breast cancer patients who underwent surgery after CNB at Seoul St. Mary's Hospital between December 2014 and December 2017. Estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki67 were analyzed using immunohistochemistry. ER and PR were evaluated by Allred score (0-8). HER2 was graded from 0 to +3, and all 2+ cases were reflex tested with silver in situ hybridization. The labeling index of Ki67 was counted by either manual scoring or digital image analysis. Molecular subtypes were classified using the above surrogate markers. In total, 629 patients were evaluated. The CRs of ER, PR, HER2, and Ki67 were 96.5% (kappa, 0.883; p<.001), 93.0% (kappa, 0.824; p<.001), 99.7% (kappa, 0.988; p<.001), and 78.7% (kappa, 0.577; p<.001), respectively. Digital image analysis of Ki67 in CNB showed better concordance with Ki67 in surgical specimens (CR, 82.3%; kappa, 0.639 for digital image analysis vs. CR, 76.2%; kappa, 0.534 for manual counting). The CRs of luminal A, luminal B, HER2, and triple negative types were 89.0%, 70.0%, 82.9%, and 77.2%, respectively. CNB was reasonably accurate for determining ER, PR, HER2, Ki67, and molecular subtypes. Using digital image analysis for Ki67 in CNB produced more accurate molecular classifications.

Digital pathology generates gigapixel whole-slide images that require extensive storage, limiting scalability and increasing costs. Traditional compression methods apply uniform ratios, disregarding variations in diagnostic importance across regions. Here we show Adaptive compression for gigapixel whole-slide images (AdaSlide), a framework that balances compression efficiency and diagnostic integrity. AdaSlide integrates a reinforcement learning-based Compression Decision Agent (CDA) that determines the optimal compression ratio per region, and a Foundational Image Enhancer (FIE) that restores visual fidelity after compression. In evaluations, pathologists achieved near-chance accuracy (55%) in distinguishing original from restored images, confirming high perceptual fidelity. Across 13 downstream tasks, AdaSlide maintained diagnostic performance in most tasks while reducing storage to 10-35% of the original size. By aligning compression strategies with clinical relevance, AdaSlide enables efficient, scalable, and reliable storage for digital pathology and future AI development.

Endometrial cancer is one of the most rapidly increasing gynaecological malignancies worldwide. The clinically adapted molecular classification of endometrial carcinoma, derived from The Cancer Genome Atlas, comprises four major subtypes: POLE-mutated, mismatch repair deficiency, p53-abnormal expression, and no specific molecular profile. Its clinical implementation has improved prognostic stratification, risk assessment, and treatment decision-making in patients with endometrial carcinoma. However, current workflows rely on immunohistochemistry and targeted sequencing, which increase costs, turnaround times, and infrastructure requirements, thereby limiting their universal adoption in routine practice. In this context, recent advances in artificial intelligence (AI), particularly deep learning models capable of predicting molecular features directly from H&E-stained whole-slide images, have emerged as promising tools for precision oncology. In addition to reproducing established molecular classification, these approaches may reveal previously unrecognised biomarker-defined histologic patterns that are difficult to detect using conventional methods. This article synthesises the current evidence on AI-based molecular classification in endometrial carcinoma from a pathologist-centred perspective, with an emphasis on the biological rationale, methodological limitations, and future directions for clinical translation.

This case describes the initial detection of cervical small cell neuroendocrine carcinoma on conventional endometrial and cervicovaginal liquid-based cytology. Given its rarity and high aggressiveness, early recognition and rapid, intensive treatment are crucial for improving patient outcomes.

Endometrial cancer is one of the most rapidly increasing gynaecological malignancies worldwide. The clinically adapted molecular classification of endometrial carcinoma, derived from The Cancer Genome Atlas, comprises four major subtypes: POLE-mutated, mismatch repair-deficient, p53-abnormal expression, and no specific molecular profile. Its clinical implementation has improved prognostic stratification, risk assessment, and treatment decision-making in patients with endometrial carcinoma. However, current workflows rely on immunohistochemistry and targeted sequencing, which increase costs, turnaround times, and infrastructure requirements, thereby limiting their universal adoption in routine clinical practice. Recent advances in artificial intelligence (AI), particularly deep learning models capable of predicting molecular features directly from H&E-stained whole-slide images, have emerged as promising tools for precision oncology. In addition to reproducing established molecular classification, these approaches may reveal previously unrecognised biomarker-defined histologic patterns that are difficult to detect using conventional methods. This article synthesises the current evidence on AI-based molecular classification in endometrial carcinoma from a pathologist-centred perspective, emphasising the biological rationale, methodological limitations, and future directions for clinical translation.

Introduction: Gastric cancer (GC) exhibits substantial molecular heterogeneity, necessitating precision oncology approaches. Conventional biomarker assessments, including immunohistochemistry, in situ hybridization, polymerase chain reaction, and next-generation sequencing, are the diagnostic standard but resource-intensive and may be limited by tissue availability, cost, and turnaround time. Artificial intelligence (AI)–enabled computational pathology using whole-slide images (WSIs) has emerged as a scalable approach for inferring molecular and immune phenotypes directly from routine hematoxylin and eosin–stained histopathological images. Methods: This review synthesizes recent advances in AI-based WSI analysis for biomarker assessment in GC, focusing on molecular subtype prediction, actionable genetic alterations, immune-related features, tumor microenvironment characterization, and multimodal integration. Results: AI models have demonstrated promising performance in predicting microsatellite instability and Epstein–Barr virus status, supporting their potential use as prescreening or triage tools to prioritize confirmatory testing. In addition, these approaches enable the quantitative characterization of the tumor microenvironment by mapping tertiary lymphoid structures and immune architecture, providing prognostic insights. Multimodal integration of histopathology with radiologic, genomic, and clinical data has shown improved predictive performance compared to single-modality approaches, particularly for recurrence and treatment responses. Discussion: However, challenges remain, including model interpretability, variability in performance across different datasets, and incomplete data across modalities. Future directions include prospective multicenter validation in clinical workflows, standardization of evaluation frameworks, and implementation of uncertainty estimation to support clinical decision-making. Overall, AI–enabled digital pathology represents a promising approach for advancing precision oncology in GC by improving biomarker assessment and providing insights into tumor biology.