Search

Synthetic data generated by generative artificial intelligence models can serve as a substitute for real patient data. In this Review, Eckardt et al. discuss how synthetic data sets can overcome barriers to data access and sharing, democratize scientific discovery in cancer research, and reduce the costs and failure rates of cancer clinical trials. They also discuss how this will only become possible if we can overcome the challenges of a lack of standardization in training data selection, model evaluation, bias mitigation, privacy preservation and quality assurance.

Five experts share their thoughts on key areas of focus in multidisciplinary cancer research for the upcoming years. They discuss the research approaches, tools, technologies, collaborations and way of thinking the lab of the future should integrate.

The authors develop multimodal machine learning models to infer metastatic recurrence risk for early-stage, hormone receptor-positive breast cancer from H&E images using >6000 cases across three centers, outperforming a nomogram and unimodal methods.

A comprehensive benchmarking of 19 histopathology foundation models finds that a vision-language foundation model outperforms vision-only models.

The accuracy of melanoma diagnosis can vary considerably among clinicians, impacting both patient outcomes and the performance of related AI tools. Here, the authors systematically assess interrater variability among expert pathologists reviewing histopathological images and clinical metadata of melanoma-suspicious lesions collected at eight German hospitals.

Ferber et al. present an autonomous artificial intelligence agent system for deployment of specialized medical oncology computational tools, validating their system across various clinical scenarios representative of typical patient care workflows.

Medical image classification remains a challenging process in deep learning. Here, the authors evaluate a large vision language foundation model (GPT-4V) with in-context learning for cancer image processing and show that such models can learn from examples and reach performance similar to specialized neural networks while reducing the gap to current state-of-the art pathology foundation models.

Molecular biomarkers of recurrence in colorectal cancer (CRC) generally cannot capture spatial information about the tumour and its microenvironment. Here, the authors develop HIBRID, a deep learning model to predict disease-free survival in CRC from histopathology whole slide images, improving risk stratification in large cohorts.

Two papers by Kather and colleagues and Gerstung and colleagues develop workflows to predict a wide range of molecular alterations from pan-cancer digital pathology slides.

Vision-language artificial intelligence models (VLMs) can be employed to recognize lesions in cancer images. Here, the authors show that VLMs can be misled by prompt injection attacks, producing harmful output and leading to incorrect diagnoses.

Artificial intelligence agents are autonomous systems that use large language models to reason and as such can perform complex, multistep tasks with minimal human oversight. This Review by Truhn et al. discusses how these agents — which have already been implemented in several industries — could transform cancer research and oncology, and looks at the challenges that need to be addressed before they can be efficiently and safely used.

The authors develop and validate mFISHseq, a spatially informed assay that tackles several unmet needs in breast cancer, including discordance in molecular subtyping and prognostic risk and identification of biomarkers predicting response to immunotherapies and antibody-drug conjugates.

An inherently explainable AI trained on 1,015 expert-annotated prostate tissue images achieved strong Gleason pattern segmentation while providing interpretable outputs and addressing interobserver variability in pathology.

Combined hepatocellular-cholangiocarcinomas (cHCC-CCA) are challenging to diagnose, as they exhibit features of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICCA). Here, the authors use deep learning to re-classify cHCC-CCA tumours into HCC or ICCA based on histopathology images.

Microsatellite instability is a known risk factor for colorectal cancer development and treatment response. Here, the authors utilise deep learning to develop MSIntuit, a pre-screening tool to detect MSI from H&E stained slides.

Generalist medical artificial intelligence (GMAI) models are gaining momentum in their applications for cancer treatment. In this Comment, Gilbert and Kather advocate for novel regulation of GMAI approaches to ensure patient safety and adequate physician support.

A comprehensive evaluation of memorization across datasets, including training samples and patient data copies, shows that latent diffusion models can memorize a diverse set of medical images with varying properties.

Artificial Intelligence can support diagnostic workflows in oncology, but they are vulnerable to adversarial attacks. Here, the authors show that convolutional neural networks are highly susceptible to white- and black-box adversarial attacks in clinically relevant classification tasks.

Artificial intelligence (AI) system is known to improve dermatologists’ diagnostic accuracy for melanoma. This group applies the eye-tracking technology on dermatologists when diagnosing dermoscopic images of melanomas and reports improved balanced diagnostic accuracy when using an X(explainable) AI system comparing to the standard one.

A decentralized, privacy-preserving machine learning framework used to train a clinically relevant AI system identifies actionable molecular alterations in patients with colorectal cancer by use of routine histopathology slides collected in real-world settings.

Cancer biomarkers are often continuous measurements, which poses challenges for their prediction using classification-based deep learning. Here, the authors develop a regression-based deep learning method to predict continuous biomarkers - such as the homologous repair deficiency score - from cancer histopathology images.

Schmatko et al. review the application of artificial intelligence to digitized histopathology for cancer diagnosis, prognosis and classification and discuss its potential utility in the clinic and broader implications for cancer research and care.

A deep residual learning framework identifies microsatellite instability in histology slides from patients with cancer and can be used to guide immunotherapy.

Technical metrics used to evaluate medical artificial intelligence tools often fail to predict their clinical impact. We characterize this discordance and propose a framework of study designs to guide the translational process for clinical artificial intelligence tools, acknowledging their diversity and specific validation requirements.

The development of clinically relevant artificial intelligence (AI) models has traditionally required access to extensive labelled datasets, which inevitably centre AI advances around large centres and private corporations. Data availability has also dictated the development of AI applications: most studies focus on common cancer types, and leave rare diseases behind. However, this paradigm is changing with the advent of foundation models, which enable the training of more powerful and robust AI systems using much smaller datasets.

Saldanha, Zhu et al. present an integrated pipeline combining weakly supervised learning with local artificial intelligence (AI) model training via swarm learning (SL) to circumvent a need for centralized data sharing. Deploying SL internationally with on-site data processing addresses challenges such as data privacy and annotation variability enabling AI training across international datasets while preserving data privacy.

The involvement of cell death pathways in the early stage of pancreatic ductal adenocarcinoma (PDAC) development, especially KRAS-dependent acinar-to-ductal metaplasia (ADM), remains to be investigated. Here, the authors find that TAK1 mediates cell survival during ADM transdifferentiation through suppression of apoptosis and necroptosis, which could be targeted for prevention and treatment of PDAC.

A knowledge gap persists between machine learning developers and clinicians. Here, the authors show that the Advanced Data Analysis extension of ChatGPT could bridge this gap and simplify complex data analyses, making them more accessible to clinicians.

Predicting disease progression is an important medical problem, but it can be challenging for end-to-end machine learning approaches. Han and colleagues demonstrate that generative models can work together with medical experts to jointly predict the progression of a disease, osteoarthritis.

Deep learning models are advancing cancer research and oncology but require human engagement to perform complex multi-step workflows. Autonomous artificial intelligence agents, empowered by large language models, present a promising solution by enabling the planning, execution and optimization of multi-step reasoning in biomedical research.

This Perspective discusses the use and potential of large language models and clinical decision support systems in gastroenterology and hepatology, highlighting opportunities, challenges and limitations of large language models and clinical decision support systems in clinical practice. Key directions for research and insights into clinical integration and safe use are also discussed.

Deep learning models have been trained on The Cancer Genome Atlas to predict numerous features directly from histology, including survival, gene expression patterns, and driver mutations. Here, the authors demonstrate that site-specific histologic signatures can lead to biased estimates of accuracy for such models, and propose a method to minimize such bias.

A deep learning model trained on multiple tumor immune cell stainings from patients with colorectal cancer outperforms currently used clinical and single tumor immune cell staining-based parameters in predicting prognosis. The model can also predict the response to neoadjuvant therapy.

Tumour evolution modelling indicates that different tumour spatial structures can determine different tumour evolutionary modes, which are regulated by cell dispersal and cell–cell interactions. Model predictions of four evolutionary modes are consistent with empirical observations of cancers with varying architectures.

To facilitate the safe and effective implementation of autonomous artificial intelligence agents in healthcare, regulatory frameworks must evolve beyond static device paradigms to incorporate adaptive oversight and flexible pathways.

Schulz et al. systematically benchmark performance scaling with increasingly sophisticated prediction algorithms and with increasing sample size in reference machine-learning and biomedical datasets. Complicated nonlinear intervariable relationships remain largely inaccessible for predicting key phenotypes from typical brain scans.

This Review provides an introductory guide to artificial intelligence (AI)-based tools for non-computational cancer researchers. Here, Perez-Lopez et al. outline the general principles of AI for image analysis, natural language processing and drug discovery, as well as how researchers can get started with each of them.

Deep learning can mine clinically useful information from histology. In gastrointestinal and liver cancer, such algorithms can predict survival and molecular alterations. Once pathology workflows are widely digitized, these methods could be used as inexpensive biomarkers. However, clinical translation requires training interdisciplinary researchers in both programming and clinical applications.