Inflammatory bowel disease (IBD) involves chronic inflammation of the digestive tract, with treatment options often burdened by adverse effects. Identifying biomarkers for personalized treatment is crucial. While immune cells play a key role in IBD, accurately identifying ulcer regions in whole slide images (WSIs) is essential for characterizing these cells and exploring potential therapeutics. Multiple instance learning (MIL) approaches have advanced WSI analysis but they lack spatial context awareness. In this work, we propose a weakly-supervised model called DomainGCN that employs a graph convolution neural network (GCN) and incorporates domain-specific knowledge of ulcer features, specifically, the presence of epithelium, lymphocytes, and debris for WSI-level ulcer prediction in IBD. We demonstrate that DomainGCN outperforms various state-of-the-art (SOTA) MIL methods and show the added value of domain knowledge.

ABSTRACT Background Fibroblast growth factor 21 (FGF21) is a metabolic regulator with demonstrated efficacy for the treatment of metabolic dysfunction‐associated steatohepatitis (MASH). FGF21 signals through ‘c’ isoforms of the FGF receptors (FGFR) 1–3 and the co‐receptor β‐klotho. Aims We report the safety and efficacy of MK‐3655, a monoclonal antibody that binds β‐klotho and selectively activates the FGFR1c/β‐klotho co‐receptor complex, in patients with pre‐cirrhotic MASH. Methods Phase 2b, randomised, multicenter, double‐blind, placebo‐controlled, parallel‐group study in patients with pre‐cirrhotic MASH (NAS ≥ 4 and MASH CRN fibrosis score Stage 2 or 3). Participants were randomised 1:1:1:1 to receive MK‐3655 50 mg, 100 mg, 300 mg, or matching placebo subcutaneously every 4 weeks. The primary endpoint was MASH resolution without worsening of fibrosis by histology at Week 52. An interim analysis (IA) of liver fat content (LFC) was planned once ≥ 25 participants per treatment group completed an MRI‐PDFF assessment at Week 24. Results Among 183 participants, mean BMI was 33.4 kg/m 2 , mean LFC was 18.1%, and 52.5% had type 2 diabetes. At the IA, the differences from placebo in relative reduction from baseline in LFC were assessed as insufficient for continuation of the trial. Among participants with Week 24 LFC assessment, percent relative reductions from baseline (LS mean difference vs. placebo) for MK‐3655 50 mg ( N = 33), 100 mg ( N = 36), and 300 mg ( N = 31), were 19.1%, 19.0%, and 26.1%, respectively. MK‐3655 was generally well tolerated. Conclusions In patients with pre‐cirrhotic MASH, treatment with MK‐3655 resulted in a modest reduction in LFC at 24 weeks. Clinical Trial Number EudraCT: 2019‐003048‐63; NCT: 04583423.

Background & Aims: Intra and inter-pathologist variability poses a significant challenge in metabolic dysfunction-associated steatohepatitis (MASH) biopsy evaluation, leading to suboptimal selection of patients and confounded assessment of histological response in clinical trials. We evaluated the utility of an artificial intelligence (AI) digital pathology (DP) platform to help pathologists improve the reliability of fibrosis staging. Methods: A total of 120 digitized histology slides from two trials (NCT03517540, NCT03912532) were analyzed by four expert hepatopathologists, with and without AI assistance in a randomized, crossover design. We utilized an AI DP platform consisting of unstained second harmonic generation/two photon excitation fluorescence (SHG/TPEF) images and AI quantitative fibrosis (qFibrosis) values. Results: AI assistance significantly improved inter-pathologist kappa for fibrosis staging, particularly for early fibrosis (F0-F2), with reduced variance around the median reads. Intra-pathologist kappa was unchanged. AI assistance increased pathologist concordance for identifying clinical trial inclusion cases (F2-F3) from 45% to 71%, exclusion cases (F0/F1/F4) from 38% to 55%, and evaluation of fibrosis response to treatment from 49% to 61%. SHG/TPEF images, qFibrosis continuous values, and qFibrosis stage were considered useful by at least three out of four pathologists in 83%, 55%, and 38% of cases, respectively. In the context of a clinical trial, the increase in inter-pathologist concordance was modeled to result in a X25% reduction in the potential need for adjudication as well as a X45% increase in the study power for a kappa improvement from X0.4 to X0.7. Conclusions: The use of AI DP enhances inter-rater reliability of fibrosis staging for MASH. This indicates that the SHG/TPEFbased AI DP tool is useful for assisting pathologists in assessing fibrosis, thereby enhancing clinical trial efficiency and reliability of fibrosis readouts in response to treatments. (c) 2024 Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA , The Author(s), HistoIndex Pte Ltd. Published by Elsevier B.V on behalf of European Association for the Study of the Liver (EASL). This is an open access article under the CC BY-NC-ND license (http:// creativecommons.org/licenses/by-nc-nd/4.0/).

4073 Background: Pathologic tumor downstaging (pDS) is more accurate than clinical staging for prognosis; generally, patients (pts) with downstaged tumors have lower recurrence rates and improved overall survival (OS). mPR is an important surrogate end point in non–small cell lung cancer, but data are scarce in gastric cancer. In the randomized phase 3 KEYNOTE-585 trial (NCT03221426), neoadjuvant or adjuvant pembro + chemo (cisplatin + capecitabine [XP] or cisplatin + 5-fluorouracil [FP]) was not superior to pbo + chemo (XP or FP) for event-free survival (EFS) in the intention-to-treat (ITT) population of pts with locally advanced G/GEJ adenocarcinoma; pembro + chemo significantly improved pathologic complete response (pCR) vs pbo + chemo (difference, 10.9%; 95% CI, 7.5-14.8; P<0.00001). In this post hoc analysis we examined the utility of mPR and pDS as surrogate end points in KEYNOTE-585. Methods: Pts with untreated localized G/GEJ adenocarcinoma (including Siewert type 2 or 3 tumors) and ECOG PS 0 or 1 who planned to undergo surgery after preoperative chemo were eligible. The main cohort of KEYNOTE-585 received pembro or pbo + XP or FP; the safety cohort received pembro or pbo + docetaxel, oxaliplatin, FP, and leucovorin (FLOT). Objectives for the post hoc analysis were the relation between mPR (≤10% residual viable tumor, comparable to Mandard tumor regression grade 1 or 2) and EFS per RECIST v1.1 per investigator and OS, plus the percentage of pts who had improved survival, by stage, compared with downstaging in the pTNM group. Overall downstaging was derived by comparing answers to questions regarding primary tumor and nodal involvement from clinical and pathologic assessments for each pt. Pts could have tumor downstaging, nodal downstaging, or both. If both, the number of levels downstaged for tumor and nodes were added together to calculate a total number of levels downstaged. The database cutoff was February 9, 2023. Results: A total of 1007 pts were analyzed (n=502, pembro + chemo; n=505, pbo + chemo); 44.0% and 34.1%, respectively, had pathologic nodal stage N0. The incidence of mPR (grade 1/2 tumor regression) was 31.5% with pembro + chemo vs 22.2% with pbo + chemo. In pts who had mPR, the HRs (95% CI) for EFS and OS were 0.6 (0.4-1.0) and 0.7 (0.4-1.2), respectively, for pembro + chemo vs pbo + chemo. Downstaging results are shown (Table). Conclusions: In this post hoc analysis, pts with mPR had numerically improved PFS and OS with pembro + chemo vs pbo + chemo; these findings were consistent with analysis of the ITT population. Further studies in this population are warranted. Clinical trial information: NCT03221426 . [Table: see text]

Human epidermal growth factor receptor 2 (HER2) serves as a prognostic and predictive biomarker for breast cancer. Recently, there has been an increasing number of studies evaluating the feasibility of utilizing H&E WSIs for determining HER2 status through innovative data-driven deep learning methods, taking advantage of the ubiquitous availability of H&E WSIs. One of the main challenges with these data-driven methods is the need for large-scale datasets with high quality annotations, which can be expensive to curate. Therefore, in this study, we explored both the region-of-interest (ROI)-based supervised and the attention-based multiple-instance-learning (MIL) weakly supervised methods for predicting HER2 status on H&E WSIs to evaluate whether avoiding labor-intensive tumor annotation will compromise the final prediction performance. The ROI-based method involved an Inception-v3 along with an aggregation step to combine the patch-level predictions into a WSI-level prediction. On the other hand, the attention-based MIL methods explored ImageNet pretrained ResNet, H&E image pretrained ResNet, and H&E image pretrained vision transformer (ViT) as encoders for WSI-level HER2 prediction. Experiments are carried out on N = 355 WSIs available in public domain with HER2 status determined by IHC and ISH and annotations of breast invasive carcinoma. The dataset was split into training/validation/test set with 80/10/10 ratio. Our results demonstrate that the attention-based ViT MIL method is able to reach similar accuracy as the ROI-based method on the independent test set (AUC of 0.79 (95% CI: 0.63-0.95) versus 0.88 (95% CI: 0.63-0.9) respectively), and thus reduces the burden of labor-intensive annotations. Furthermore, the attention mechanism enhances interpretability of the results and offers insights into the reliability of the predictions.

H&E images can be utilized to predict genetic mutations as biomarkers to potentially substitute many molecular biomarker assays in order to aid patients. Having a single model built by conducting prediction tasks simultaneously can save computation resources and provide a more generalizable model for future usage. A basic technique for generating such a comprehensive and efficient model is to employ a multi-task learning approach. However, overfitting the model to the trivial answers can occur in training for multiple tasks with extremely imbalanced class labels where resampling and rebalancing for all minor classes simultaneously are prohibited. Herein we propose a sequential multi-task learning approach to train a single model capable of predicting multiple genetic mutations while avoiding overfitting to trivial answers for imbalanced classes. We compared our strategy to the baseline multi-task training, as well as two more advanced approaches: (1) using weighted loss and (2) using self-supervised pre-training. We also used a trimming method to deal with noisy labels. To assess our methods, we trained models to predict 10 genetic mutations on the H&E images of the TCGA-LUAD dataset. AUROC and F1 score are reported, while we demonstrate that F1 score may be a more suitable metric for multi-task learning with imbalanced labels. It is shown that our proposed trimming strategy combined with sequential learning could improve the predictions on all of the mutations compared with other multi-task learning approaches. Also, we investigated the application of continual learning.

Tumor mutation burden (TMB) is an important biomarker for the prediction of response to anti-PD-1 immunotherapies. Studies have shown that higher level of TMB (TMB-H) is associated with higher response rate to immunotherapies in patients with various types of advanced solid tumors. However, the measurement of TMB depends on whole exome sequencing (WES) which is an expensive assay and not always available in standard clinical oncology settings. In this work, we assess the feasibility of predicting TMB-H based upon hematoxylin and eosin (H&E)-stained histopathology images, which is a routinely conducted assay in clinical oncology. Using an Inception-V3 convolutional neural network (CNN) as a baseline feature extractor, we compare adding a multi-layer perceptron (MLP) and a squeeze-and-excitation (SE) network on top of the baseline CNN. Training from random initialization and tuning with pretrained weights are also compared. Experiments are conducted on the H&E whole-slide images (WSI) of the melanoma dataset of The Cancer Genome Atlas (TCGA). Results from a 4-fold cross-validation show that the highest average area under the receiver operating characteristic curve (AUC) is 0.589, which implies that the prediction of TMB based on H&E WSI for melanoma remains a challenging problem that will warrant further investigations.

The PD-L1 IHC 22C3 pharmDx used on the Dako Autostainer Link 48 (ASL48) staining platform is an established method for assessing programmed death-ligand 1 (PD-L1) expression in tumor tissue and determining patient eligibility for pembrolizumab treatment; however, the availability of this platform is limited in Europe and Asia. The aims of this study were to develop and optimize protocols for the PD-L1 22C3 antibody concentrate with multiple immunohistochemistry staining platforms and to validate these protocols using PD-L1 combined positive score (CPS) with a cut-off of ≥ 1 in gastric or gastroesophageal junction adenocarcinoma. The 22C3 antibody concentrate was tested and optimized protocols were developed for use with three staining platforms: Dako ASL48, Ventana BenchMark ULTRA, and Leica BOND-MAX. Tumor specimens (N = 120) from patients with gastric or gastroesophageal junction adenocarcinoma were used for the validation study; these specimens were evaluated independently by three pathologists for PD-L1 CPS as a continuous variable and using a cut-off of ≥ 1. PD-L1 IHC 22C3 pharmDx used on the Dako ASL48 platform served as the reference or gold standard. The intraclass correlation coefficient of CPS as a continuous variable between the gold standard and each staining platform assessed was 0.910–0.989. When CPS was dichotomized based on a cut-off of ≥ 1, depending on the pathologist and the platform used, positive percentage agreement was 81–99

We give quasipolynomial‐time approximation algorithms for designing networks with a minimum degree. Using our methods, one can design networks whose connectivity is specified by “proper” functions, a class of 0–1 functions indicating the number of edges crossing each cut. We also provide quasipolynomial‐time approximation algorithms for finding two‐edge‐connected spanning subgraphs of approximately minimum degree of a given two‐edge‐connected graph, and a spanning tree (branching) of approximately minimum degree of a directed graph. The degree of the output network in all cases is guaranteed to be at most (1 + ϵ) times the optimal degree, plus an additive O (log 1+ϵ n ) for any ϵ > 0. Our analysis indicates that the degree of an optimal subgraph for each of the problems above is well estimated by certain polynomially solvable linear programs. This suggests that the linear programs we describe could be useful in obtaining optimal solutions via branch and bound. © 2004 Wiley Periodicals, Inc. NETWORKS, Vol. 44(3), 203–215 2004

Consider a directed graph G=V,E) with n vertices and a root vertex r∈V. The DMDST problem for G is one of constructing a spanning tree rooted at r, whose maximal degree is the smallest among all such spanning trees. The problem is known to be NP-hard. A quasipolynomial time approximation algorithm for this problem is presented. The algorithm finds a spanning tree whose maximal degree is at most O(Δ*) + log n) where, Δ* is the degree of some optimal tree for the problem. The running time of the algorithm is shown to be O n log n log n. Experimental results are presented showing that the actual running time of the algorithm is much smaller in practice.