Pathologists diagnose and grade prostate cancer using thin 2-dimensional (2D) histologic sections, but these 3 to 5 micron sections are too thin to visualize complete glandular networks and 3-dimensional (3D) spatial relationships of adenocarcinomas. We hypothesized that understanding volumetric glandular organization would reveal architectural features associated with prostate cancer progression and biochemical recurrence (BCR). We analyzed 2 archived prostatectomy cohorts using different sampling methods: simulated 1 mm core-needle biopsies from the University of Washington and 3 × 1 mm-punch biopsies from the University of Pennsylvania. We used open-top light-sheet microscopy to visualize intact tissue networks and developed GlaSkeN, a computational pathology framework to quantify 3D prostatic gland architecture. GlaSkeN used deep learning to segment glandular structures from 3D images and then constructed skeleton-based representations to extract volumetric features, including branch length, branching angles, torsion, and curvature. We analyzed associations between architectural features and 5-year BCR-free survival using 6-fold cross-validated Cox regression. GlaSkeN identified 3D architectural features significantly associated with BCR in both cohorts: the University of Washington (hazard rati [HR], 5.18; 95% CI, 1.18-22.68; C-index = 0.68; P = .019) and the University of Pennsylvania (HR, 2.04; 95% CI, 1.14-3.65; C-index = 0.62; P < .05). In multivariable analysis, GlaSkeN remained prognostic after controlling for clinicopathological variables (HR, 2.30; 95% CI, 1.13-4.7; P = .021). Limitations include different sampling methods between cohorts and limited sample sizes. This 3D analysis captured glandular organization, spatial connectivity, and branching patterns unassessable in 2D cross-sections. GlaSkeN identified glandular architecture features associated with BCR independent of standard clinical variables, suggesting 3D architecture could provide additional prognostic information to complement current histopathological grading. Validation in larger independent cohorts is warranted.

Recent breakthroughs in artificial intelligence through foundation models and agents have accelerated the evolution of computational pathology. Demonstrated performance gains reported across academia in benchmarking datasets in predictive tasks such as diagnosis, prognosis, and treatment response have ignited substantial enthusiasm for clinical application. Despite this development momentum, real world adoption has lagged, as implementation faces economic, technical, and administrative challenges. Beyond existing discussions of technical architectures and comparative performance, this review considers how these emerging AI systems can be responsibly integrated into medical practice by connecting deployable clinical relevance with downstream analytical capabilities and their technical maturity, operational readiness, and economic and regulatory context. Drawing on perspectives from an international group, we provide a practical assessment of current capabilities and barriers to adoption in patient care settings.

Hirschsprung disease (HD) is a congenital disorder characterized by the absence of ganglion cells in the colonic nervous plexuses, resulting in bowel obstruction and various complications. The diagnosis of HD demands expertise, experience, and ancillary tests. Our objective was to design an artificial intelligence (AI) tool to facilitate the diagnostic process of HD and evaluate its performance in terms of sensitivity, specificity, and average area under the curve (AUC). Using a camera-equipped microscope, we digitized 222 high-power fields of view obtained from H&E-stained slides. Nuclei were segmented using a deep learning algorithm and manually annotated by an expert pathologist. From these, 2076 nuclei were selected, including 346 ganglion cells and 1730 non-ganglion cells. A set of 100 features related to shape, color, and texture was extracted from the nuclei. To evaluate their utility in distinguishing the two categories, a cross-validation scheme was employed. The nuclei were randomly divided into training (70%) and validation sets (30%). The Wilcoxon signed-rank test was employed to identify the top features in the training set, which were then used to train an AI classifier to distinguish between the two categories: Ganglion cell and non-ganglion cell. The classifier's performance was assessed using the validation set, yielding an average AUC = 0.98. The classifier's performance was assessed using the validation set, yielding an average AUC = 0.98. Taken together, our findings indicate that this AI-driven framework may serve as a valuable support tool for pathologists, facilitating the diagnosis of HD in routine clinical practice.

PURPOSE:Trastuzumab-based chemotherapy has improved outcomes in human epidermal growth factor receptor 2 (HER2)-positive breast cancer, but treatment benefit varies among patients. Predictive signatures are needed to identify patients most likely to respond to these therapies. EXPERIMENTAL DESIGN:We developed Density and Spatial architecture of Tumor-Infiltrating Lymphocytes (DeSTIL), a computational signature derived from hematoxylin and eosin slides. The signature captures spatial organization of immune cells and interactions with nonimmune cells. DeSTIL was trained on HER2+ breast cancer slides from The Cancer Genome Atlas (n = 250) and validated in a phase III National Surgical Adjuvant Breast and Bowel Project (NSABP) B-41 randomized clinical trial (n = 221), which compared chemotherapy plus trastuzumab, lapatinib, or combination. The DeSTIL scores were dichotomized into positive and negative groups, and event-free survival (EFS) was assessed using Cox proportional hazards with interaction terms. RESULTS:In NSABP B-41, DeSTIL-positive patients (n = 61) showed significantly improved event-free survival (EFS) with trastuzumab compared with the combination arm [hazard ratio (HR) = 0.09; 95% confidence interval (CI) = 0.01-0.77; P = 0.006] and a significant signature-treatment interaction (P = 0.024). No EFS difference was observed in DeSTIL-negative patients (n = 160). Gene expression analysis supported the image-derived signature stratifying DeSTIL-positive and DeSTIL-negative tumors. In an exploratory pathologic complete response analysis, a classifier trained on University Hospitals Cleveland slides achieved AUCs of 0.70 in the training cohort and 0.63 in the trastuzumab arm of the NSABP B-41 validation cohort. CONCLUSIONS:DeSTIL identifies a subset of HER2+ patients who derive greater benefit from trastuzumab. These findings support the potential of computationally derived immune architecture to inform selection of standard HER2-targeted therapies.

Abstract Background: Multiple myeloma (MM) disproportionately affects men (57% vs. 43% women), yet the biological basis for these sex differences remains unexplored. The bone-marrow immune microenvironment, particularly plasma-immune spatial organisation, plays a critical role in disease progression and therapeutic response. Emerging evidence suggests that sex-specific immunologic differences contribute to disparities in cancer biology and outcomes, yet these patterns remain poorly characterised in MM. Understanding whether plasma-immune architectural features differ by sex and whether they carry prognostic value may inform more personalised risk stratification in MM. Methods: We identified 104 MM whole-slide bone-marrow biopsies from the Cancer Moonshot Biobank. Using pre-trained deep learning models, immune cells were segmented, followed by plasma cell detection. Morphological (density and spatial) patterns of plasma cells and other immune cells were quantified in peri-tumoral and non-tumoral compartments, and FDR-corrected Welch's t-tests were performed to compare gender specific differences. Prognostic associations of these features were assessed using univariate Cox proportional hazard models. Results: On comparing the immune cell phenotypes, males exhibited higher plasma density relative to tumor density (0.158 vs female 0.105, FDR p=0.029). Spatial analysis showed that males also had higher plasma-lymphocyte cluster overlap in the peri-tumoral stroma (0.418 vs female 0.238, FDR p= 0.036). In survival analysis, overlap of plasma-lymphocyte cluster was prognostic of overall survival: HR = 1.730 (95% CI: 1.074-2.787), p= 0.018, c-index = 0.552. In contrast, there were no significant differences in density or spatial features of lymphocytes across the genders. Conclusions: Plasma cell shows gender specific differences in density and spatial arrangement, despite no significant differences in lymphocyte morphology. These sex-specific patterns appear to predict survival, supporting the need for sex-stratified risk assessment and personalised prognostication strategies in MM. Citation Format: Dharini Raghavan, Advait Madabhushi, Amritpal Singh, Tilak Pathak, GERMAN CORREDOR, Ajay K. Nooka, Anant Madabhushi. Sex-specific plasma-immune architectural differences in bone marrow predicts overall survival in multiple myeloma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 5078.

3118 Background: Improved risk stratification is critically needed across multiple cancers to guide treatment decisions. This includes reducing overtreatment in Ductal Carcinoma in Situ (DCIS), addressing late recurrence risk in estrogen receptor-positive/lymph node-negative (ER+/LN-) breast cancer (BC), and refining prognosis in head & neck squamous cell carcinomas (HNSCC), and high-grade serous ovarian carcinoma (HGSOC). Tumor-infiltrating lymphocytes (TILs) reflect anti-tumor immunity, whereas tumor-associated macrophages (TAMs) polarize toward anti-tumor M1-like (CD68+) or immunosuppressive M2-like (CD163+) phenotypes. Integrating TILs with TAM polarization markers may better capture inflammatory balance and improve outcome prediction. We developed a hematoxylin and eosin (H&E)-only AI method (VAIP) to quantify TILs and virtual CD68/CD163 TAMs to develop multi-cancer based potential prognostic models. Methods: TILs were detected on H&E with HoVer-UNet, and virtual CD68 and CD163 macrophages were inferred with VISTA. In cancer regions, we computed cell densities and immune-balance ratios and trained cancer-specific Cox models with compact feature sets (DCIS 6; HGSOC 7; ER+ 4; HNSCC 4). Prognostic stratification used disease-free survival (DFS) for DCIS (N=273), HGSOC (training N=165, independent test N=220), and ER+ BC (training N=144, test N=622), and overall survival (OS) for HNSCC (N=77), reflecting clinical endpoints. Evaluation used cross-validation for cohorts without an external test set (DCIS, HNSCC) and independent training and testing when independent cohort was available (HGSOC, ER+). Results: Risk stratification performance across cancer types is in Table 1. The DCIS model used 6 features, largely TIL-macrophage contrast ratios (TIL/M2, TIL/M1). HGSOC used 7 features with a similar ratio-heavy immune-balance signature, plus macrophage-normalized density and polarization metrics (e.g., M1/N, M/N). In contrast, ER+ BC and HNSCC achieved best performance with 4 features dominated by macrophage measures, particularly M2 density and the M1/M2 ratio. Across sites, CD163-related macrophage metrics and immune-balance ratios were consistently informative. Conclusions: Integrating H&E-visible TILs with virtual CD68/CD163 TAMs enables fully automated, H&E-only prognostic biomarkers that stratify OS and DFS across multiple cancers, demonstrating consistent prognostic value in DCIS, ER+ BC, HGSOC, and HNSCC. These findings support further study of this histomorphometric risk classifier through broader external and prospective validation. Performance across cancer types. Cancer Type Cohort C-index HR [95% CI] p-value DCIS All 0.616 1.76 [1.20-2.59] <0.001 DCIS RT/No-treatment 0.624 1.94 [1.15-3.26] 0.003 HGSOC Test 0.553 1.38 [1.02-1.87] 0.030 ER+ BC Test 0.577 1.51 [1.00-2.28] 0.004 HNSCC All 0.603 1.92 [1.03-3.36] 0.012

1122 Background: Pathological complete response (pCR) remains difficult to predict in triple-negative breast cancer (TNBC) patients treated with immune checkpoint inhibitors due to limited validated biomarkers. While tumor-infiltrating lymphocytes (TILs) are prognostic in TNBC, they are commonly treated as a composite population. Plasma cells are central to humoral immunity and can influence antitumor response through antibody production, antigen presentation, and T-cell modulation. However, the independent and cooperative prognostic role of plasma cells, either alone or in spatial interaction with other immune cells has not been systematically evaluated, particularly in TNBC. Here, we evaluated the value of plasma cells to predict pathological complete response using AI-based analysis of routine H&E slides. Methods: We analyzed 239 stage I–III TNBC patients across four cohorts. Models were trained on KEYNOTE-522 regimen treated patients (D1, N=130) and validated in both the Yale TNBC immunotherapy trial cohort (D2, N=65; NCT02489448) treated with durvalumab (anti-PD-L1 antibody) and University Hospital cohort (D3, N=25) treated with pembrolizumab. Plasma cells and lymphocytes were automatically detected on whole-slide H&E images using deep learning–based nuclear segmentation and tumor region classification. Quantified metrics included immune cell density, tumor-infiltrating plasma cell (TIP) ratios, and spatial features capturing plasma cell clustering and interaction with lymphocytes. Density-based classifiers were trained on D1 and evaluated for pCR prediction. Results: Plasma cell density alone predicted pCR with higher accuracy than lymphocyte density across both validation cohorts. In D2), plasma cell density achieved an AUC of 73.2 compared with 62.9 for lymphocyte density. Similarly, in the external institutional cohort (D3), plasma cell density outperformed lymphocyte density (AUC 69.2 vs 60.8, respectively; Table 1). These findings demonstrate the superior predictive value of plasma cell density over conventional lymphocyte-based measures for pCR in immunotherapy-treated TNBC. Conclusions: Plasma cells demonstrate independent and cooperative predictive value in immunotherapy-treated TNBC. AI-derived plasma cell density and spatial interaction features improve pCR prediction and survival stratification beyond conventional lymphocyte-based metrics. Comparison of plasma cell performance compared to lymphocytes on holdout set D2 (Yale TNBC clinical trial NCT02489448) and D3. Holdout 1 Holdout 2 Yale NCT02489448 (N=64), D2 UH Hospital (N=25), D3 M plasma 73.2 69.2 M lymph 62.9 60.8

Supplementary Figures S1-S11 and Supplementary Tables S1-S4

First-line treatment for small-cell lung cancer (SCLC) involves platinum-based chemotherapy and immunotherapy for extensive (ES) and limited (LM) disease, respectively. Rapid progression and metastasis highlight the need for improved biomarkers. We developed PhenopyCell, a computational pathology tool that quantifies immune-tumor spatial architecture on Hematoxylin and Eosin (H&E) slides to predict outcomes. Developing PhenopyCell for spatial quantification of immune-tumor interactions and clinical outcome prediction. Retrospective study of 281 SCLC patients (149 LM, 132 ES) treated with platinum chemotherapy (2010-2020) from multi-institutional archives, divided into training (D1, n = 101) and validation (D2/D3, n = 180) cohorts. PhenopyCell extracted 101 spatial features (immune clustering, tumor density) from whole-slide images. Overall survival (OS) via Cox models and chemotherapy response through ROC and precision-recall analyses. PhenopyCell-derived features correlated with OS and treatment response across datasets (D1 HR = 1.66, P = 0.036; D2 HR = 1.98, P = 0.04; D3 HR = 2.13, P = 0.04). Stratified analyses showed strong prognostic value for both ES-SCLC (HR up to 5.11) and LM-SCLC (HR up to 34.91). Chemotherapy response prediction achieved AUCs of 0.62-0.79. PhenopyCell independently predicts survival and therapy response in SCLC, outperforming conventional histopathology and supporting personalized treatment approaches.

6073 Background: Laryngeal Squamous Cell Carcinoma (LaSCC) has varying outcomes based on the stage of cancer patients present with. Currently, a high proportion of patients are diagnosed with advanced-stage LaSCC complicating the treatment landscape and over-treating low risk patients. Risk stratification of LaSCC can help tailor treatment plans. Numerous studies have identified spatial architecture of tumor-infiltrating lymphocytes (TILs) as a prognostic biomarker in oral cavity and oropharyngeal SCC. In this work, we evaluate the prognostic value of an artificial intelligence (AI)-leveraged approach that characterizes the spatial architecture of TILs on digitized hematoxylin and eosin (H&E)-stained slides from patients with LaSCC. Methods: H&E slides from 192 patients with LaSCC were collected from Baylor Medical Center. This dataset was randomly divided into two equal cohorts, A and B. The slides were digitized as whole slide images at 40x magnification. The nuclei of all cells were automatically segmented using a deep-learning model (Hover-Net). Each nucleus was then classified as TIL or non-TIL based on morphological features. TILs and non-TILs were clustered based on proximity, and features related to the density and spatial distribution were extracted. The top features, determined by the least absolute shrinkage and selection operator, were used to train a Cox Proportional Hazards regression model that assigned a risk score for recurrence of cancer to each patient in cohort A. For validation, the model was applied to patients in cohort B. The 25 th percentile training risk score was used as a cutoff for classifying patients as high or low risk. The performance of the model in prognosticating loco-regional recurrence (LRR) was evaluated using survival analysis. Results: Patients in Cohort B identified as “high risk” by the model based on spatial organization of TILs had a significantly shorter survival time. Univariate survival analysis showed this model was prognostic for DFS with a hazard ratio of 2.57 (95% Confidence Interval: 1.12-5.89, p-value=0.048), meaning that patients classified as “high risk” are approximately 2.5 times more likely to develop LLR. Conclusions: We used computational pathology to characterize the architecture of TILs and develop a model to predict risk of LLR in LaSCC. With additional validation, this approach can be used to assist clinicians with making clinical decisions. Multivariate analysis. Variable Reference vs Comparison Pr(>|z|) HR (95% CI) Race Black vs Caucasian 0.93 0.92 (0.40 - 2.31) N N0 vs N+ 0.67 1.80 (0.12 - 26.06) T 1-2 vs 3-4 0.83 1.28 (0.13 - 12.38) Chemo Yes vs No 0.38 3.49 (0.21 - 56.90) Tobacco Yes vs No 0.26 2.08 (0.58 - 7.47) Alcohol Yes vs No 0.43 1.53 (0.54 - 4.35) TIL Architecture Risk High vs Low 0.03 * 0.26 (0.08 - 0.89)

Accurate prognosis prediction is essential for guiding cancer treatment and improving patient outcomes. While recent studies have demonstrated the potential of histopathological images in survival analysis, existing models are typically developed in a cancer-specific manner, lack extensive external validation, and often rely on molecular data that are not routinely available in clinical practice. To address these limitations, we present PROGPATH, a unified model capable of integrating histopathological image features with routinely collected clinical variables to achieve pancancer prognosis prediction. PROGPATH employs a weakly supervised deep learning architecture built upon the foundation model for image encoding. Morphological features are aggregated through an attention-guided multiple instance learning module and fused with clinical information via a cross-attention transformer. A router-based classification strategy further refines the prediction performance. PROGPATH was trained on 7999 whole-slide images (WSIs) from 6,670 patients across 15 cancer types, and extensively validated on 17 external cohorts with a total of 7374 WSIs from 4441 patients, covering 12 cancer types from 8 consortia and institutions across three continents. PROGPATH achieved consistently superior performance compared with state-of-the-art multimodal prognosis prediction models. It demonstrated strong generalizability across cancer types and robustness in stratified subgroups, including early- and advanced-stage patients, treatment cohorts (radiotherapy and pharmaceutical therapy), and biomarker-defined subsets. We further provide model interpretability by identifying pathological patterns critical to PROGPATH’s risk predictions, such as the degree of cell differentiation and extent of necrosis. Together, these results highlight the potential of PROGPATH to support pancancer outcome prediction and inform personalized cancer management strategies.

INTRODUCTION:Measuring the chromatin state of a tumor provides a powerful map of its epigenetic commitments; however, as these are generally bulk measurements, it has not yet been possible to connect changes in chromatin accessibility to the pathological signatures of complex tumors. In parallel, recent advances in computational pathology have enabled the identification of spatial features and immune cells within oral cavity tumors and their microenvironment. METHODS:Here, we present pathogenomic fingerprinting (PaGeFin), a novel method that integrates morphological tumor features with chromatin states using ATAC-seq. This framework links spatial morphologic and epigenetic features, offering insights into tumor progression and immune evasion within and across tumors. Morphologic features describing spatial relationships between tumor and lymphocyte cells that are prognostic of oral cavity squamous cell carcinoma (OSCC) were identified through AI-driven pathology analysis. These pathomic features were spatially colocalized within the epigenome of 4 distinct sections of 4 OSCC tumors. RESULTS:These key features pinpointed chromatin regions responsible for critical immune cell function through peak locations and enrichment analysis, highlighting loci of CD27+ memory B cells, helper CD4+ T cells, and cytotoxic CD8 naïve T cells that likely drive morphologic changes in the distribution of lymphocytes in the tumor microenvironment and promote aggressive tumor behavior. Gene Ontology analysis revealed that the CTLA4, CD79A, CD3D, and CCR7 genes were embedded in these regions. CONCLUSION:This computational approach is the first to assess the correlation between pathomic and epigenetic features in the context of cancer.

We developed a computational pathology pipeline to extract and analyze collagen disorder architecture (CoDA) features from whole slide images (WSIs) of 2,212 colon cancer (CC) patients across multiple institutions. CoDA features—capturing collagen fragmentation, bundling, anisotropy, density, and rigidity, were evaluated for associations with clinical variables (overall stage, T/N/M stage), molecular classifications (Consensus Molecular Subtypes [CMS1–4]), and genetic mutations (KRAS, BRAF, NRAS) using the Mann-Whitney U test with Bonferroni correction. These analyses revealed significant differences in CoDA feature distributions across multiple subgroups, suggesting that collagen architecture varies meaningfully with tumor stage, molecular subtype, and mutation status.To assess how well CoDA features could distinguish between these subgroups, we implemented a Random Forest classification framework. High mean AUC values (≥0.7) across several variables indicated strong discriminatory performance of CoDA features in separating clinically and biologically distinct groups. For survival analysis, LASSO-Cox models were trained on the PLCO dataset to generate CoDA-based risk scores for overall survival (OS) and disease-free survival (DFS), which were used to stratify patients into high- and low-risk groups in a combined validation dataset (TCGA, UH, and Emory). Kaplan-Meier curves demonstrated significant survival differences across clinical stages, CMS subtypes, and KRAS mutation status. Multivariable Cox proportional hazards models further confirmed the independent prognostic value of CoDA features after adjusting for clinical, molecular, and genetic covariates. These findings highlight that CoDA features are significantly associated with key clinical and molecular characteristics, can distinguish relevant patient subgroups, and offer independent prognostic information, underscoring their potential utility in characterizing the tumor microenvironment and informing risk stratification in CC.