Segmentation Model Benchmarking: A Strategic Prerequisite for Robust Geospatial Foundation Models

Introduction: Tumor single-cell data provide unique insights into the cellular dynamics of cancer progression and therapeutic outcomes. Inspired by advancements in natural language processing (NLP), foundation models are increasingly used in single-cell analyses to uncover biological insights and identify clinical biomarkers. However, the performance of these models on cancer-specific tasks remains underexplored. We present an extensive evaluation of foundation models trained on millions of single-cell profiles for cancer-focused investigations. Methods: We evaluated nine predictive models-three traditional methods and six foundation models (four Geneformer and two scGPT models), across various downstream tasks. Traditional models used embeddings from high variable genes (HVG), principal components (PCA), and generative model embedding (scVI), each paired with a random forest classifier for downstream classification. The evaluated foundation models varied in architecture sizes, pretraining dataset sizes (30M to 95M cell profiles), and training datasets, which were either generic or cancer-specific. We assessed the quality of representations by measuring their predictive performance on lung cancer treatment status (TKI treated vs. treatment-naive patients) and breast cancer subtypes (ER+ vs. TNBC). We examined how model characteristics influenced their performance in these domains for all tasks. Results: Zero-shot evaluations of the foundation models showed superior single-cell data clustering, measured by average silhouette width (ASW), compared to traditional methods (HVG, PCA, scVI), suggesting that pre-trained embeddings from millions of single-cell profiles may offer enhanced representations. We thus evaluated the learned representation on downstream predictive tasks, showing that foundation models with a larger pretraining dataset (95M cells), cancer-specific training, larger input size (4096 genes), and deeper model (12 layers) achieved the best performance among foundation models (F1=0.97) for predicting treatment status in lung cancer. In contrast, smaller and generic foundation models had significantly lower performance, however, additional self-supervised training on task-specific data rescued their performance to match larger foundation models. Similar trends were observed in predicting breast cancer subtypes. Notably, the best foundation model’s performance was not significantly different from the best traditional method (PCA) in predicting breast cancer subtypes. Conclusions: These results highlight key factors driving foundation model performance on cancer-specific tasks. While current foundation models show inconsistent contributions to cancer-specific questions, observed performance trends highlight their potential to enhance single-cell data representation and drive significant advancements in cancer research and precision medicine. Citation Format: Haitham Elmarakeby, Ahmed Roman, Shreya Johri, Eliezer M. Van Allen. Evaluation of single-cell foundation models for cancer outcome predictions [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5084.

Study question Can a generalizable computer vision foundation model be trained using self-supervised learning on time-lapse embryo images to perform multiple clinically relevant downstream tasks? Summary answer We developed FEMI, a foundation model trained on eight million time-lapse images, to perform multiple clinical tasks, including blastocyst quality scoring, ploidy prediction, and segmentation. What is known already In vitro fertilization success critically depends on choosing viable embryos, a process hampered by current limited diagnostic tools, high costs, and ethical concerns. In recent years, various medical fields have increasingly explored foundation models using vision transformer architectures. These models, trained in a self-supervised manner on vast unlabeled image datasets, perform various clinically relevant tasks. Once trained on a massive unlabeled dataset, the encoder portion of the foundation model can be extracted and subsequently fine-tuned with a labeled dataset to perform various clinically relevant tasks like classification, regression, and segmentation. Study design, size, duration The foundation model was trained on 8 million Embryoscope® (E-SD) and Embryoscope+® (E+) images spanning clinics in the United States, Canada, and Europe. Datasets contained various additional information, such as ploidy status, blastocyst scores (numerical values from 3 to 14 based on Zhan et al. 2020), and segmentation masks (for trophectoderm [TE], inner cell mass [ICM], and zona pellucida [ZP]) that were used for training downstream tasks. Participants/materials, setting, methods A masked autoencoder architecture is trained on time-lapse images to create a foundation model (FEMI). The encoder from the autoencoder is extracted and fine-tuned on three image-based downstream tasks: ploidy prediction, blastocyst scoring, and embryo component segmentation. The performance of the fine-tuned foundation model is compared to VGG16 architectures, specifically trained for each downstream task. We evaluated performances using the area under the receiver-operating-characteristic (AUROC), mean absolute error (MAE), and mean intersection over union (mIoU). Main results and the role of chance The training dataset for the downstream tasks consisted of image data and labels, including PGT-A ploidy results (euploid or aneuploid), blastocyst scores (3-14), and segmentation masks. No clinical data like maternal age was used in order to assess FEMI’s learning capabilities from only time-lapse images against VGG16 model baselines. Both FEMI and baseline models used the same data splits for consistent comparisons. For ploidy prediction, models trained on a Weill Cornell Medicine (WCM) dataset were validated using WCM, Florida, and Spain data. FEMI achieved a 0.610 ± 0.004 AUROC, outperforming the baseline’s 0.590 ± 0.005. In blastocyst score prediction, FEMI attained a superior MAE of 0.099 ± 0.002 compared to the baseline’s 0.118 ± 0.001. For embryo segmentation (TE, ICM, ZP), a publicly available dataset from Simon Fraser University was split for training and validation. FEMI exceeded the baseline in mIoU for TE segmentation (0.738 ± 0.002 vs. 0.726 ± 0.002) and was comparable in ICM (0.832 ± 0.008 vs. 0.821 ± 0.011) and ZP segmentation (0.779 ± 0.002 vs. 0.778 ± 0.010). Limitations, reasons for caution The current version of FEMI is trained on a subset of our available datasets and can be improved through further training. Moreover, the downstream ploidy and blastocyst score models may be biased by errors in PGT results and the subjectivity of embryologists, respectively. Wider implications of the findings A generalizable foundational model for IVF time-lapse imaging aids clinicians in embryo selection by providing various clinical insights for better decision-making. FEMI, once fully trained, will be publicly accessible to the scientific community allowing for researchers to fine-tune FEMI on their own clinic-specific applications with their own image datasets. Trial registration number not applicable

Foundation models (FMs) have revolutionized generative AI (GAI) lifecycle with their pre-trained intelligence capabilities. While the recent success of web-based models like GPT-4 has spurred interest in extending FMs to edge scenarios like the Internet of Vehicles (IoV), challenges such as data privacy, network congestion, and limited edge resource awareness hinder the direct application of cloud-based FMs. To address these challenges, this paper reconfigures the mixture of experts (MoE) architecture to deploy Transformer-based FM experts across vehicles for distributed GAI inference. Specifically, we propose an MoE-empowered vehicular FM system with two key innovations: a physical gating network that dynamically adapts to wireless environments, and a vehicle-to-vehicle (V2V) communication-based expert parallelism mechanism to enhance efficiency and resource utilization. Then, we formulate the communication, computation, and memory models for analyzing inference latency. Furthermore, to optimize the computational resource allocation for enhancing inference performance, we establish a Stackelberg game and propose the Gradient Ascent and Evolutionary Optimization-based Competitive Pricing and Allocation (GECPA) algorithm, which balances resource allocation and usage costs by combining the rapid convergence of gradient ascent with the broader exploration capability of evolutionary optimization. Simulation results demonstrate the superior parallel processing efficiency and reduced latency of the proposed MoE-based FM inference scheme. Compared with the best-performing benchmark algorithm, GECPA improves the average utility of infrastructure vehicles by up to 2.01 times, reduces inference latency by 16.58%, and increases the successful execution rate of GAI tasks by 10.66%, thereby achieving a better balance between efficiency and incentive compatibility in dynamic IoV environments.

Vision foundation models in remote sensing have been extensively studied due to their superior generalization on various downstream tasks. Synthetic Aperture Radar (SAR) offers all-day, all-weather imaging capabilities, providing significant advantages for Earth observation. However, establishing a foundation model for SAR image interpretation inevitably encounters the challenges of insufficient information utilization and poor interpretability. In this paper, we propose a remote sensing foundation model based on complex-valued SAR data, which simulates the polarimetric decomposition process for pre-training, i.e., characterizing pixel scattering intensity as a weighted combination of scattering bases and scattering coefficients, thereby endowing the foundation model with physical interpretability. Specifically, we construct a series of scattering queries, each representing an independent and meaningful scattering basis, which interact with SAR features in the scattering query decoder and output the corresponding scattering coefficient. To guide the pre-training process, polarimetric decomposition loss and power self-supervised loss are constructed. The former aligns the predicted coefficients with Yamaguchi coefficients, while the latter reconstructs power from the predicted coefficients and compares it to the input image's power. The performance of our foundation model is validated on nine typical downstream tasks, achieving state-of-the-art results. Notably, the foundation model can extract stable feature representations and exhibits strong generalization, even in data-scarce conditions.

The prevalence of the pre-training & fine-tuning paradigm enables machine learning models to quickly adapt to various downstream tasks by fine-tuning pre-trained foundation models (FMs), greatly facilitating various IoT applications that rely on model inference in dynamic edge serverless environments. Efficiently dispatching inference requests and configuring instances to batch inference requests can significantly enhance resource efficiency. However, existing serverless inference solutions are tailored for traditional models, make coarse-grained request dispatching and instance configuration decisions, fail to exploit the shared model backbone characteristics of the FM and capture delayed rewards in dynamic environments, and ignore communication latency between edge sites, resulting in high costs and constraint violations. In this paper, we leverage our insight that fine-grained batch inference requests can effectively exploit the shared model backbone feature of FM to save monetary costs. We propose an algorithm that incorporates deep reinforcement learning (DRL) and expert intervention for fine-grained request dispatching and instance configuration, where the DRL component outputs fractional solutions as guidance, while the expert intervention module integrates our insights—batching reduces monetary costs at the expense of increased inference latency, whereas higher configurations shorten inference latency. This module rounds fractional solutions and adjusts instance configurations to search for optimal solutions while satisfying constraints, with theoretical guarantees rigorously proved. Finally, we conducted our experiments on an OpenFaas-based platform and simulator, and extensive trace-driven evaluation results show that ExpertDRL can save costs by up to 85.14% and improve request acceptance ratio by up to 26.93%, compared to the state-of-the-art solution.

The foundation model has recently garnered significant attention due to its potential to revolutionize the field of visual representation learning in a self-supervised manner. While most foundation models are tailored to effectively process RGB images for various visual tasks, there is a noticeable gap in research focused on spectral data, which offers valuable information for scene understanding, especially in remote sensing (RS) applications. To fill this gap, we created for the first time a universal RS foundation model, named SpectralGPT, which is purpose-built to handle spectral RS images using a novel 3D generative pretrained transformer (GPT). Compared to existing foundation models, SpectralGPT 1) accommodates input images with varying sizes, resolutions, time series, and regions in a progressive training fashion, enabling full utilization of extensive RS Big Data; 2) leverages 3D token generation for spatial-spectral coupling; 3) captures spectrally sequential patterns via multi-target reconstruction; and 4) trains on one million spectral RS images, yielding models with over 600 million parameters. Our evaluation highlights significant performance improvements with pretrained SpectralGPT models, signifying substantial potential in advancing spectral RS Big Data applications within the field of geoscience across four downstream tasks: single/multi-label scene classification, semantic segmentation, and change detection.

Artificial Intelligence (AI) is transforming science, industry, and daily life at an unprecedented pace. Foundation Models (FMs) now shape how we search, create, and interact with information across domains from finance to healthcare. Yet, paradoxically, most AI research produced in academic and industrial labs remains confined to benchmark performance and paper metrics. The rapid increase in computational resources and in the size of the research community has contributed to this progress by fueling an unprecedented surge in publications, with venues such as the Conference on Neural Information Processing Systems (NeurIPS) exhibiting an annual growth rate of approximately 29% since 2017. This apparent progress is reflected in the continual stream of reported state-of-the-art (SOTA) results for the majority of publications. Yet, despite the vast number of proposed methods and the accompanying media attention, only a small fraction translate into real-world impact. This discrepancy raises a fundamental question: how can methodological research better align with the goal of practical translation? This thesis addresses overlooked elements of the model development pipeline that critically influence whether research outputs achieve translation. We investigate tasks situated upstream of model training, such as labeling instructions and internal Quality Assurance (QA), as well as downstream tasks, including user-centric benchmarking. These questions are particularly urgent in healthcare, where methodological weaknesses can translate into risks for patient safety. To illustrate this challenge, we focus on surgery as a representative high-stakes domain. Across these studies, we demonstrate that design choices in these stages can and often exert a greater impact on model performance and applicability than many architectural innovations. Building on our analysis, we propose new methods and best practices for sustainable research and reliable translation. First, we demonstrate that the majority of biomedical imaging projects does not provide any labeling instructions, even though annotators consistently consider them crucial. In the largest study on labeling instructions to date, to our knowledge, we demonstrate that visual examples are key to improving annotation quality, whereas solely extending textual descriptions yields no measurable or even negative benefit. To address these shortcomings, we provide best practices for involving annotators in both the annotation process and the construction of labeling instructions. Following a comprehensive review by the Medical Image Computing and Computer Assisted Intervention Society (MICCAI) Special Interest Group for Biomedical Image Analysis Challenges, future challenges are now required to adhere to stricter standards, including the mandatory publication of labeling instructions. Second, we turn our focus to internal QA, which at the time represented a standard practice in annotation workflows for professional annotation companies. We demonstrate that this additional step yields only marginal improvements in annotation quality. In resource-constrained scenarios, we further show that investment in labeling instructions provides greater benefits than additional QA. To guide more effective QA, we developed a statistical model and identified generalizable image characteristics for which QA is most beneficial. This ultimately enables users to selectively apply QA to images with a high likelihood of improvement instead of processing every image. Third, we show that with the growing number of benchmarks, domain users must carefully select their Vision-Language Model (VLM) of choice. To support users with their domain-specific benchmarking, we developed DomainBench, a framework for turning user-selected image datasets into diverse and scalable VLM benchmarks. To help the community adopt our framework, we released seven new datasets covering a wide range of domains from kitchen environments to animal scenes. To complement the generation process, we introduced a new metric, Accuracy%(t), that accounts for shared base images across different tasks, which previous work neglects. Lastly, we apply DomainBench to the surgical domain, where robust translation of AI methods is especially critical, and provide the first comprehensive assessment of VLMs in surgical imaging by benchmarking 15 SOTA VLMs across more than 150,000 question-answer pairs. We demonstrate that current VLMs can handle elementary surgical perception tasks but struggle on questions that require medically informed reasoning. We further provide a critical comparison of medical versus generalist VLMs, uncovering that specialized medical VLMs are outperformed by generalist VLMs. Our findings emphasize the need for future work on knowledge integration strategies that extend beyond conventional finetuning. Together, these contributions highlight that translation in Machine Learning (ML) depends less on incremental architectural advances than on careful design of the model development pipeline. By addressing labeling instructions, internal QA, user-centric benchmarking, and domain-specific validation, this thesis provides both methodological insights and practical frameworks that strengthen the reliability and applicability of ML models. In doing so, it contributes to aligning methodological progress with real-world impact across natural and medical imaging and ultimately improving the scientific standards in the field.

With the advent of large language models (LLMs), managing scientific literature via LLMs has become a promising direction of research. However, existing approaches often overlook the rich structural and semantic relevance among scientific literature, limiting their ability to discern the relationships between pieces of scientific knowledge, and suffer from various types of hallucinations. These methods also focus narrowly on individual downstream tasks, limiting their applicability across use cases. We propose LitFM, the first literature foundation model designed for a wide variety of practical downstream tasks on domain-specific literature, with a focus on citation information. At its core, LitFM contains a novel graph retriever that can provide accurate and diverse recommendations for LLM to integrate graph structure information and relevant literature. LitFM also leverages a knowledge-infused LLM, fine-tuned through a well-developed instruction paradigm. It enables LitFM to extract domain-specific knowledge from literature and reason relationships among them. By integrating citation graphs during both training and inference, LitFM can generalize to unseen papers and accurately assess their relevance within existing literature. Additionally, we introduce new large-scale literature citation benchmark datasets on three academic fields, featuring sentence-level citation information and local context. Extensive experiments validate the superiority of LitFM, achieving 28.1% improvement on retrieval task in precision, and an average improvement of 7.52% over state-of-the-art across six downstream literature-related tasks.

A foundation model, a large-scale deep learning model pretrained on vast datasets, has revolutionized data interpretation through self-supervised learning with capacity for various downstream tasks. Concurrently, single-cell genomics is in urgent need of unified frameworks capable of integrating and comprehensively analyzing the rapidly expanding data repositories. Inspired by advances in foundation models, researchers have extended these techniques to single-cell analysis, giving rise to single-cell foundation models (scFMs). Typically, these models use transformer architectures to incorporate diverse omics data and extract latent patterns at both cell and gene/feature levels for the analysis of cellular heterogeneity and complex regulatory networks. Despite their promise, scFMs face challenges including the nonsequential nature of omics data, inconsistency in data quality and the computational intensity required for training and fine-tuning. Furthermore, interpreting the biological relevance of latent embeddings and model representations remains nontrivial. Here we provide an overview of scFMs, highlighting their key concepts and applications across downstream tasks. We critically assess the current limitations and propose future directions aimed at enhancing the robustness, interpretability and scalability of scFMs. Ultimately, addressing these challenges will be crucial for establishing scFMs as pivotal tools in advancing single-cell genomics and unlocking deeper insights into cellular function and disease mechanisms.

Accurate mapping of weed infestations in fallow farmlands is critical for supporting sustainable weed management and reducing unnecessary chemical inputs. Previous work has demonstrated the effectiveness of convolutional encoder–decoder architectures, such as U-Net, for weed segmentation from satellite imagery; however, these approaches are typically constrained by sensor-specific training, limited cross-site generalisation, and sensitivity to variations in spectral and spatial resolution. In this study, we investigate the application of the ANYSAT foundation model [1] for sensor-agnostic weed segmentation across heterogeneous fallow agricultural farms across Australia. Building on top of an established U-Net-based workflow, we evaluate whether a foundation model pretrained on diverse Earth observation data can improve robustness and transferability across multiple satellite sensors without explicit sensor-dependent retraining. Multi-spectral satellite imagery from different platforms is used to fine-tune ANYSAT for semantic segmentation of weed presence in fallow paddocks, with human-curated and U-Net-refined weed masks serving as supervisory labels. We design a systematic evaluation strategy based on leave-one-farm and leave-one-region validation to test model robustness under spatial and spectral variability. Rather than focusing on achieved performance, this work emphasises assessing feasibility, identifying the strengths and limitations of foundation-model-based segmentation for this task, and outlining key considerations for operational deployment in data-sparse agricultural settings. By framing weed detection as a sensor-agnostic problem, this study provides a structured pathway for testing foundation models in agroecosystem monitoring. It contributes to understanding how emerging Earth observation foundation models can be adapted for practical agricultural applications. [1] Astruc, Guillaume, et al. "AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.

RNA biology is undergoing a transformative revolution driven by AI foundation models. These models learn the intricate relationships between RNA sequence, structure, and function by training on vast, diverse datasets spanning millions of RNA molecules across various species. Through self-supervised learning on these sequences, these models acquire a generalizable understanding of RNA, which can then be fine-tuned for various downstream tasks, thereby enabling the decoding of functional rules embedded in RNA sequences. In this review, we provide a comprehensive guide to RNA foundation models. Using concrete examples of RNA biology, we begin with the concept of foundation models and review the importance of pre-training datasets, architectural innovations, self-supervised strategies, and fine-tuning approaches that allow general RNA representations to be translated into task-specific models. Crucially, we highlight how explainable AI (XAI) methods transform these models from black-box predictors into valuable discovery tools that reveal candidate cis-regulatory elements and structural motifs. As RNA foundation models keep advancing and integrating more multimodal biological data, they aim to uncover additional regulatory rules and functions encoded in RNA.

Earth observation (EO) repositories comprise Petabytes of data. Due to their widespread use, these repositories experience extremely large volumes of data transfers. For example, users of the Sentinel Data Access System downloaded 78.6 PiB of data in 2022 alone. The transfer of such data volumes between data producers and consumers causes substantial latency and requires significant amounts of energy and vast storage capacities. This work introduces Neural Embedding Compression (NEC), a method that transmits compressed embeddings to users instead of raw data, greatly reducing transfer and storage costs. The approach uses general purpose embeddings from Foundation Models (FM), which can serve multiple downstream tasks and neural compression, which balances between compression rate and the utility of the embeddings. We implemented the method by updating a minor portion of the FM’s parameters (approximately 10%) for a short training period of about 1% of the original pre-training iterations. NEC’s effectiveness is assessed through two EO tasks: scene classification and semantic segmentation. When compared to traditional compression methods applied to raw data, NEC maintains similar accuracy levels while reducing data by 75% to 90%. Notably, even with a compression rate of 99.7%, there’s only a 5% decrease in accuracy for scene classification. In summary, NEC offers a resource-efficient yet effective solution for multi-task EO modeling with minimal transfer of data volumes.

Earth observation (EO) yields large-scale, multimodal datasets collected from various satellite missions, DEMs, land-use data, and textual metadata. Foundation models like 4M (Massively Multimodal Masked Modeling) can learn a joint embedding space that bridges modality gaps, mitigates missing data issues, and facilitates partial spatio-temporal alignment [1]. However, directly training such foundation models on the vast, high-dimensional original EO datasets is not only computationally intensive but also imposes substantial demands on storage resources.To address this, one can leverage VQ-VAE (Vector Quantized-Variational AutoEncoder) as neural compressors to transform high-dimension multimodal inputs into a few discrete indices, significantly reducing data volume while preserving critical information. By inverting the tokenization process, we can reconstruct the original high-dimensional data with minimal quality loss, aided by adversarial and perceptual losses that enhance reconstruction fidelity.Traditional VQ-based approaches, however, face challenges such as inefficient codebook utilization and limited latent space representation. To overcome these, we propose scaling strategies that complement 4M’s tokenizer-based architecture. By expanding the codebook size, latent dimensions, and network depth, our method captures the complexity of EO modalities more effectively. Specifically, we employ spherical quantization techniques like Grouped Spherical Quantization (GSQ) to address limitations in traditional approaches [2]. GSQ constrains codebook vectors to a spherical surface, stabilizing training, preventing code collapse, and promoting uniform codebook usage. Unlike standard VQ, GSQ uses spherical initialization and normalization to maintain consistent distances among codebook entries, ensuring robust latent space coverage even under extreme compression or large codebooks. From our empirical and ablation studies, alternative methods like LFQ (Lookup-Free Quantization), FSQ (Finite Scalar Quantization), and RVQ (Residual Vector Quantizer) often exhibit limitations, such as tightly coupling the latent dimension to codebook size or relying on specialized training losses. In contrast, spherical-based techniques effectively decouple latent dimensions from codebook vocabulary, providing greater flexibility and scalability as data demands increase.Our approach enables neural compressors to adapt to varying scales of compression and complexity without compromising performance. Comprehensive scalability experiments—examining large codebooks, deeper networks, and diverse compression ratios—assessed the generalizability of the proposed compression strategies and demonstrated their effectiveness on high-dimensional, large-scale EO data with minimal information loss. By integrating advanced compression techniques with scalable architectures, this framework establishes a robust foundation for addressing multimodal challenges in EO research that significantly reduces the difficulty of training foundation models on multimodal high-dimensional EO data.

Learning a robust foundation model against clean-label data poisoning attacks at downstream tasks

BackgroundThe assessment of cortical thickness through structural magnetic resonance imaging (MRI) reveals various cortical atrophy patterns, delivering crucial insights into the extent of neurodegeneration. Numerous studies have focused on these thickness values to construct models for diverse neurodegeneration‐related tasks, such as classifying disease groups, staging progression, or predicting progression speed. However, these tasks require substantial training datasets tailored to each specific task, necessitating the development of new task‐specific models. We aimed to establish a foundational model using pre‐trained BERT, integrating context‐aware deep latent representations derived from a large cohort of unlabeled cortical atrophy data, to address these challenges.MethodWe collected 10,450 T1‐weighted MRI images within the ADNI dataset. Cortical thickness was derived using FreeSurfer across 68 regions based on the Desikan‐Killiany atlas. Emplying the BERT model, we applied its self‐attention capabilities to encode deep bidrectional representations of the data. The model is pre‐trained with 8,360 unlabeled cortical thickness data through self‐supervised learning, predicting randomly masked thickness values. Subsequently, we fine‐tuned this pre‐trained model with only one additional output layer using a supervised approach, employing independent data subsets for three challenging downstream classification tasks: (1) mild cognitive impairment (MCI) and Alzheimer’s dementia (AD), (2) MCI converters (C‐MCI) and non‐converters (N‐MCI) within 3 years, and (3) Aβ‐positive and negative cases. During fine‐tuning, demographic information encoded by one layer was also incorporated for classification purposes.ResultTable 1 provides demographic details of the subset data for fine‐tuning, showing comparable trends in each set of information, except for MMSE in the MCI and AD comparison. Table 2 outlines our model's performance on three downstream classification tasks. Despite training with a relatively limited amount of data, our model demonstrated competitive performance with state‐of‐the‐art methods, achieving 85% accuracy on the classification of MCI and AD, 85% of C‐MCI and N‐MCI, and 69% of Aβ‐positive and negative.ConclusionThis study introduces the novel cortical atrophy foundation model, demonstrating competitive performance even with a limited dataset across diverse challenging tasks. The model's performance in predictive tasks highlights its potential applicability for various nerodegeneration‐related challenges, paving the way for advanced AI applications within the realm of neurodegenerative diseases.