RESUMO
Artificial intelligence (AI) solutions that automatically extract information from digital histology images have shown great promise for improving pathological diagnosis. Prior to routine use, it is important to evaluate their predictive performance and obtain regulatory approval. This assessment requires appropriate test datasets. However, compiling such datasets is challenging and specific recommendations are missing. A committee of various stakeholders, including commercial AI developers, pathologists, and researchers, discussed key aspects and conducted extensive literature reviews on test datasets in pathology. Here, we summarize the results and derive general recommendations on compiling test datasets. We address several questions: Which and how many images are needed? How to deal with low-prevalence subsets? How can potential bias be detected? How should datasets be reported? What are the regulatory requirements in different countries? The recommendations are intended to help AI developers demonstrate the utility of their products and to help pathologists and regulatory agencies verify reported performance measures. Further research is needed to formulate criteria for sufficiently representative test datasets so that AI solutions can operate with less user intervention and better support diagnostic workflows in the future.
Assuntos
Inteligência Artificial , Patologia , Humanos , Previsões , Conjuntos de Dados como AssuntoRESUMO
BACKGROUND: Tumor-Adipose-Feature (TAF) as well as SARIFA (Stroma AReactive Invasion Front Areas) are two histologic features/biomarkers linking tumor-associated adipocytes to poor outcomes in colorectal cancer (CRC) patients. Whereas TAF was identified by deep learning (DL) algorithms, SARIFA was established as a human-observed histopathologic biomarker. METHODS: To study the overlap between TAF and SARIFA, we performed a systematic pathological review of TAF based on all published image tiles. Additionally, we analyzed the presence/absence of TAF in SARIFA-negative CRC cases to elucidate the biologic and prognostic role of a direct tumor-adipocyte contact. TCGA-CRC gene expression data is investigated to assess the association of FABP4 (fatty-acid binding protein 4) and CD36 (fatty-acid translocase) with both TAF and CRC prognosis. RESULTS: By investigating the TAF/SARIFA overlap, we show that many TAF patches correspond to the recently described SARIFA-phenomenon. Even though there is a pronounced morphological and biological overlap, there are differences in the concepts. The presence of TAF in SARIFA-negative CRCs is not associated with poor outcomes in this cohort, potentially highlighting the importance of a direct tumor-adipocyte interaction. Upregulation of FABP4 and CD36 gene expression seem both linked to a poor prognosis in CRC. CONCLUSIONS: By proving the substantial overlap between human-observed SARIFA and DL-based TAF as morphologic biomarkers, we demonstrate that linking DL-based image features to independently developed histopathologic biomarkers is a promising tool in the identification of clinically and biologically meaningful biomarkers. Adipocyte-tumor-cell interactions seem to be crucial in CRC, which should be considered as biomarkers for further investigations.
Different methods exist in assessing samples removed from cancer patients during surgery. We linked two independently established tissue-based methods for determining the outcome of colorectal cancer patients together: tumor adipose feature (TAF) and Stroma AReactive Invasion Front Areas (SARIFA). SARIFA as biological feature was observed solely by humans and TAF was identified by the help of a computer algorithm. We examined TAF in many cancer slides and looked at whether they showed similarities to SARIFA. TAF often matched SARIFA, but not always. Interestingly, these methods could be used to predict outcomes for patients and are associated with specific gene expression involved in tumor and fat cell interaction. Our study shows that combining computer algorithms with human expertize in evaluating tissue samples can identify meaningful features in patient samples, which may help to predict the best treatment options.
RESUMO
Over the past decade, artificial intelligence (AI) methods in pathology have advanced substantially. However, integration into routine clinical practice has been slow due to numerous challenges, including technical and regulatory hurdles in translating research results into clinical diagnostic products and the lack of standardized interfaces. The open and vendor-neutral EMPAIA initiative addresses these challenges. Here, we provide an overview of EMPAIA's achievements and lessons learned. EMPAIA integrates various stakeholders of the pathology AI ecosystem, i.e., pathologists, computer scientists, and industry. In close collaboration, we developed technical interoperability standards, recommendations for AI testing and product development, and explainability methods. We implemented the modular and open-source EMPAIA Platform and successfully integrated 14 AI-based image analysis apps from eight different vendors, demonstrating how different apps can use a single standardized interface. We prioritized requirements and evaluated the use of AI in real clinical settings with 14 different pathology laboratories in Europe and Asia. In addition to technical developments, we created a forum for all stakeholders to share information and experiences on digital pathology and AI. Commercial, clinical, and academic stakeholders can now adopt EMPAIA's common open-source interfaces, providing a unique opportunity for large-scale standardization and streamlining of processes. Further efforts are needed to effectively and broadly establish AI assistance in routine laboratory use. To this end, a sustainable infrastructure, the non-profit association EMPAIA International, has been established to continue standardization and support broad implementation and advocacy for an AI-assisted digital pathology future.
RESUMO
Open and practical exchange, dissemination, and reuse of specimens and data have become a fundamental requirement for life sciences research. The quality of the data obtained and thus the findings and knowledge derived is thus significantly influenced by the quality of the samples, the experimental methods, and the data analysis. Therefore, a comprehensive and precise documentation of the pre-analytical conditions, the analytical procedures, and the data processing are essential to be able to assess the validity of the research results. With the increasing importance of the exchange, reuse, and sharing of data and samples, procedures are required that enable cross-organizational documentation, traceability, and non-repudiation. At present, this information on the provenance of samples and data is mostly either sparse, incomplete, or incoherent. Since there is no uniform framework, this information is usually only provided within the organization and not interoperably. At the same time, the collection and sharing of biological and environmental specimens increasingly require definition and documentation of benefit sharing and compliance to regulatory requirements rather than consideration of pure scientific needs. In this publication, we present an ongoing standardization effort to provide trustworthy machine-actionable documentation of the data lineage and specimens. We would like to invite experts from the biotechnology and biomedical fields to further contribute to the standard.
RESUMO
AI development in biotechnology relies on high-quality data to train and validate algorithms. The FAIR principles (Findable, Accessible, Interoperable, and Reusable) and regulatory frameworks such as the In Vitro Diagnostic Regulation (IVDR) and the Medical Device Regulation (MDR) specify requirements on specimen and data provenance to ensure the quality and traceability of data used in AI development. In this paper, a framework is presented for recording and publishing provenance information to meet these requirements. The framework is based on the use of standardized models and protocols, such as the W3C PROV model and the ISO 23494 series, to capture and record provenance information at various stages of the data generation and analysis process. The framework and use case illustrate the role of provenance information in supporting the development of high-quality AI algorithms in biotechnology. Finally, the principles of the framework are illustrated in a simple computational pathology use case, showing how specimen and data provenance can be used in the development and documentation of an AI algorithm. The use case demonstrates the importance of managing and integrating distributed provenance information and highlights the complex task of considering factors such as semantic interoperability, confidentiality, and the verification of authenticity and integrity.
Assuntos
Algoritmos , Biotecnologia , Inteligência ArtificialRESUMO
Access to large volumes of so-called whole-slide images-high-resolution scans of complete pathological slides-has become a cornerstone of the development of novel artificial intelligence methods in pathology for diagnostic use, education/training of pathologists, and research. Nevertheless, a methodology based on risk analysis for evaluating the privacy risks associated with sharing such imaging data and applying the principle "as open as possible and as closed as necessary" is still lacking. In this article, we develop a model for privacy risk analysis for whole-slide images which focuses primarily on identity disclosure attacks, as these are the most important from a regulatory perspective. We introduce a taxonomy of whole-slide images with respect to privacy risks and mathematical model for risk assessment and design . Based on this risk assessment model and the taxonomy, we conduct a series of experiments to demonstrate the risks using real-world imaging data. Finally, we develop guidelines for risk assessment and recommendations for low-risk sharing of whole-slide image data.
Assuntos
Inteligência Artificial , Privacidade , Processamento de Imagem Assistida por Computador/métodos , Diagnóstico por Imagem/métodosRESUMO
The current move towards digital pathology enables pathologists to use artificial intelligence (AI)-based computer programmes for the advanced analysis of whole slide images. However, currently, the best-performing AI algorithms for image analysis are deemed black boxes since it remains - even to their developers - often unclear why the algorithm delivered a particular result. Especially in medicine, a better understanding of algorithmic decisions is essential to avoid mistakes and adverse effects on patients. This review article aims to provide medical experts with insights on the issue of explainability in digital pathology. A short introduction to the relevant underlying core concepts of machine learning shall nurture the reader's understanding of why explainability is a specific issue in this field. Addressing this issue of explainability, the rapidly evolving research field of explainable AI (XAI) has developed many techniques and methods to make black-box machine-learning systems more transparent. These XAI methods are a first step towards making black-box AI systems understandable by humans. However, we argue that an explanation interface must complement these explainable models to make their results useful to human stakeholders and achieve a high level of causability, i.e. a high level of causal understanding by the user. This is especially relevant in the medical field since explainability and causability play a crucial role also for compliance with regulatory requirements. We conclude by promoting the need for novel user interfaces for AI applications in pathology, which enable contextual understanding and allow the medical expert to ask interactive 'what-if'-questions. In pathology, such user interfaces will not only be important to achieve a high level of causability. They will also be crucial for keeping the human-in-the-loop and bringing medical experts' experience and conceptual knowledge to AI processes.
Assuntos
Inteligência Artificial , Patologistas , Humanos , Algoritmos , Processamento de Imagem Assistida por ComputadorRESUMO
Data quality has recently become a critical topic for the research community. European guidelines recommend that scientific data should be made FAIR: findable, accessible, interoperable and reusable. However, as FAIR guidelines do not specify how the stated principles should be implemented, it might not be straightforward for researchers to know how actually to make their data FAIR. This can prevent life-science researchers from sharing their datasets and pipelines, ultimately hindering the progress of research. To address this difficulty, we developed the BIBBOX, which is a platform that supports researchers publishing their datasets and the associated software in a FAIR manner.
Assuntos
Aplicativos MóveisRESUMO
Importance: Identifying new prognostic features in colon cancer has the potential to refine histopathologic review and inform patient care. Although prognostic artificial intelligence systems have recently demonstrated significant risk stratification for several cancer types, studies have not yet shown that the machine learning-derived features associated with these prognostic artificial intelligence systems are both interpretable and usable by pathologists. Objective: To evaluate whether pathologist scoring of a histopathologic feature previously identified by machine learning is associated with survival among patients with colon cancer. Design, Setting, and Participants: This prognostic study used deidentified, archived colorectal cancer cases from January 2013 to December 2015 from the University of Milano-Bicocca. All available histologic slides from 258 consecutive colon adenocarcinoma cases were reviewed from December 2021 to February 2022 by 2 pathologists, who conducted semiquantitative scoring for tumor adipose feature (TAF), which was previously identified via a prognostic deep learning model developed with an independent colorectal cancer cohort. Main Outcomes and Measures: Prognostic value of TAF for overall survival and disease-specific survival as measured by univariable and multivariable regression analyses. Interpathologist agreement in TAF scoring was also evaluated. Results: A total of 258 colon adenocarcinoma histopathologic cases from 258 patients (138 men [53%]; median age, 67 years [IQR, 65-81 years]) with stage II (n = 119) or stage III (n = 139) cancer were included. Tumor adipose feature was identified in 120 cases (widespread in 63 cases, multifocal in 31, and unifocal in 26). For overall survival analysis after adjustment for tumor stage, TAF was independently prognostic in 2 ways: TAF as a binary feature (presence vs absence: hazard ratio [HR] for presence of TAF, 1.55 [95% CI, 1.07-2.25]; P = .02) and TAF as a semiquantitative categorical feature (HR for widespread TAF, 1.87 [95% CI, 1.23-2.85]; P = .004). Interpathologist agreement for widespread TAF vs lower categories (absent, unifocal, or multifocal) was 90%, corresponding to a κ metric at this threshold of 0.69 (95% CI, 0.58-0.80). Conclusions and Relevance: In this prognostic study, pathologists were able to learn and reproducibly score for TAF, providing significant risk stratification on this independent data set. Although additional work is warranted to understand the biological significance of this feature and to establish broadly reproducible TAF scoring, this work represents the first validation to date of human expert learning from machine learning in pathology. Specifically, this validation demonstrates that a computationally identified histologic feature can represent a human-identifiable, prognostic feature with the potential for integration into pathology practice.
Assuntos
Adenocarcinoma , Neoplasias do Colo , Masculino , Humanos , Idoso , Neoplasias do Colo/diagnóstico , Patologistas , Inteligência Artificial , Aprendizado de Máquina , Medição de RiscoRESUMO
Studies have shown that colorectal cancer prognosis can be predicted by deep learning-based analysis of histological tissue sections of the primary tumor. So far, this has been achieved using a binary prediction. Survival curves might contain more detailed information and thus enable a more fine-grained risk prediction. Therefore, we established survival curve-based CRC survival predictors and benchmarked them against standard binary survival predictors, comparing their performance extensively on the clinical high and low risk subsets of one internal and three external cohorts. Survival curve-based risk prediction achieved a very similar risk stratification to binary risk prediction for this task. Exchanging other components of the pipeline, namely input tissue and feature extractor, had largely identical effects on model performance independently of the type of risk prediction. An ensemble of all survival curve-based models exhibited a more robust performance, as did a similar ensemble based on binary risk prediction. Patients could be further stratified within clinical risk groups. However, performance still varied across cohorts, indicating limited generalization of all investigated image analysis pipelines, whereas models using clinical data performed robustly on all cohorts.
RESUMO
BACKGROUND: Presence of lymph node metastasis (LNM) influences prognosis and clinical decision-making in colorectal cancer. However, detection of LNM is variable and depends on a number of external factors. Deep learning has shown success in computational pathology, but has struggled to boost performance when combined with known predictors. METHODS: Machine-learned features are created by clustering deep learning embeddings of small patches of tumor in colorectal cancer via k-means, and then selecting the top clusters that add predictive value to a logistic regression model when combined with known baseline clinicopathological variables. We then analyze performance of logistic regression models trained with and without these machine-learned features in combination with the baseline variables. RESULTS: The machine-learned extracted features provide independent signal for the presence of LNM (AUROC: 0.638, 95% CI: [0.590, 0.683]). Furthermore, the machine-learned features add predictive value to the set of 6 clinicopathologic variables in an external validation set (likelihood ratio test, p < 0.00032; AUROC: 0.740, 95% CI: [0.701, 0.780]). A model incorporating these features can also further risk-stratify patients with and without identified metastasis (p < 0.001 for both stage II and stage III). CONCLUSION: This work demonstrates an effective approach to combine deep learning with established clinicopathologic factors in order to identify independently informative features associated with LNM. Further work building on these specific results may have important impact in prognostication and therapeutic decision making for LNM. Additionally, this general computational approach may prove useful in other contexts.
When colorectal cancers spread to the lymph nodes, it can indicate a poorer prognosis. However, detecting lymph node metastasis (spread) can be difficult and depends on a number of factors such as how samples are taken and processed. Here, we show that machine learning, which involves computer software learning from patterns in data, can predict lymph node metastasis in patients with colorectal cancer from the microscopic appearance of their primary tumor and the clinical characteristics of the patients. We also show that the same approach can predict patient survival. With further work, our approach may help clinicians to inform patients about their prognosis and decide on appropriate treatments.
RESUMO
BACKGROUND: Artificial Intelligence (AI) is becoming more and more important especially in datacentric fields, such as biomedical research and biobanking. However, AI does not only offer advantages and promising benefits, but brings about also ethical risks and perils. In recent years, there has been growing interest in AI ethics, as reflected by a huge number of (scientific) literature dealing with the topic of AI ethics. The main objectives of this review are: (1) to provide an overview about important (upcoming) AI ethics regulations and international recommendations as well as available AI ethics tools and frameworks relevant to biomedical research, (2) to identify what AI ethics can learn from findings in ethics of traditional biomedical research - in particular looking at ethics in the domain of biobanking, and (3) to provide an overview about the main research questions in the field of AI ethics in biomedical research. METHODS: We adopted a modified thematic review approach focused on understanding AI ethics aspects relevant to biomedical research. For this review, four scientific literature databases at the cross-section of medical, technical, and ethics science literature were queried: PubMed, BMC Medical Ethics, IEEE Xplore, and Google Scholar. In addition, a grey literature search was conducted to identify current trends in legislation and standardization. RESULTS: More than 2,500 potentially relevant publications were retrieved through the initial search and 57 documents were included in the final review. The review found many documents describing high-level principles of AI ethics, and some publications describing approaches for making AI ethics more actionable and bridging the principles-to-practice gap. Also, some ongoing regulatory and standardization initiatives related to AI ethics were identified. It was found that ethical aspects of AI implementation in biobanks are often like those in biomedical research, for example with regards to handling big data or tackling informed consent. The review revealed current 'hot' topics in AI ethics related to biomedical research. Furthermore, several published tools and methods aiming to support practical implementation of AI ethics, as well as tools and frameworks specifically addressing complete and transparent reporting of biomedical studies involving AI are described in the review results. CONCLUSIONS: The review results provide a practically useful overview of research strands as well as regulations, guidelines, and tools regarding AI ethics in biomedical research. Furthermore, the review results show the need for an ethical-mindful and balanced approach to AI in biomedical research, and specifically reveal the need for AI ethics research focused on understanding and resolving practical problems arising from the use of AI in science and society.
Assuntos
Inteligência Artificial , Pesquisa Biomédica , Bancos de Espécimes Biológicos , Big Data , Consentimento Livre e EsclarecidoRESUMO
Provenance is information describing the lineage of an object, such as a dataset or biological material. Since these objects can be passed between organizations, each organization can document only parts of the objects life cycle. As a result, interconnection of distributed provenance parts forms distributed provenance chains. Dependant on the actual provenance content, complete provenance chains can provide traceability and contribute to reproducibility and FAIRness of research objects. In this paper, we define a lightweight provenance model based on W3C PROV that enables generation of distributed provenance chains in complex, multi-organizational environments. The application of the model is demonstrated with a use case spanning several steps of a real-world research pipeline - starting with the acquisition of a specimen, its processing and storage, histological examination, and the generation/collection of associated data (images, annotations, clinical data), ending with training an AI model for the detection of tumor in the images. The proposed model has become an open conceptual foundation of the currently developed ISO 23494 standard on provenance for biotechnology domain.
RESUMO
The process of finding a diagnosis in the medical domain relies on implicit knowledge and the experience of a human expert. In this article, we report on the observation of human decision making, shown by the example of pathology. By tracking the diagnostic steps, individual building blocks are identified, which not only contribute to a diagnostic finding, but can also be used in the future to train and develop artificial intelligence (AI) algorithms. This work also provides insights into the interaction of human experts regarding the observation time of so-called "hot spots," the magnification used for specific findings, and the overall observation and decision path followed. The documentation scheme yields a standardized examination procedure that shows the concept the pathologist is actually looking for as well as the possible features of findings that can be identified. This contribution indicates how important visualization is for human-centered AI, and specifically for enabling human oversight with respect to AI implementation in high-stake areas, such as medicine.
Assuntos
Algoritmos , Inteligência Artificial , Humanos , Resolução de Problemas , Tomada de DecisõesRESUMO
The distributed nature of modern research emphasizes the importance of collecting and sharing the history of digital and physical material, to improve the reproducibility of experiments and the quality and reusability of results. Yet, the application of the current methodologies to record provenance information is largely scattered, leading to silos of provenance information at different granularities. To tackle this fragmentation, we developed the Common Provenance Model, a set of guidelines for the generation of interoperable provenance information, and to allow the reconstruction and the navigation of a continuous provenance chain. This work presents the first version of the model, available online, based on the W3C PROV Data Model and the Provenance Composition pattern.
Assuntos
Disciplinas das Ciências Biológicas , Reprodutibilidade dos TestesRESUMO
Artificial Intelligence (AI) for the biomedical domain is gaining significant interest and holds considerable potential for the future of healthcare, particularly also in the context of in vitro diagnostics. The European In Vitro Diagnostic Medical Device Regulation (IVDR) explicitly includes software in its requirements. This poses major challenges for In Vitro Diagnostic devices (IVDs) that involve Machine Learning (ML) algorithms for data analysis and decision support. This can increase the difficulty of applying some of the most successful ML and Deep Learning (DL) methods to the biomedical domain, just by missing the required explanatory components from the manufacturers. In this context, trustworthy AI has to empower biomedical professionals to take responsibility for their decision-making, which clearly raises the need for explainable AI methods. Explainable AI, such as layer-wise relevance propagation, can help in highlighting the relevant parts of inputs to, and representations in, a neural network that caused a result and visualize these relevant parts. In the same way that usability encompasses measurements for the quality of use, the concept of causability encompasses measurements for the quality of explanations produced by explainable AI methods. This paper describes both concepts and gives examples of how explainability and causability are essential in order to demonstrate scientific validity as well as analytical and clinical performance for future AI-based IVDs.
Assuntos
Inteligência Artificial , Redes Neurais de Computação , Algoritmos , Aprendizado de Máquina , SoftwareRESUMO
Deriving interpretable prognostic features from deep-learning-based prognostic histopathology models remains a challenge. In this study, we developed a deep learning system (DLS) for predicting disease-specific survival for stage II and III colorectal cancer using 3652 cases (27,300 slides). When evaluated on two validation datasets containing 1239 cases (9340 slides) and 738 cases (7140 slides), respectively, the DLS achieved a 5-year disease-specific survival AUC of 0.70 (95% CI: 0.66-0.73) and 0.69 (95% CI: 0.64-0.72), and added significant predictive value to a set of nine clinicopathologic features. To interpret the DLS, we explored the ability of different human-interpretable features to explain the variance in DLS scores. We observed that clinicopathologic features such as T-category, N-category, and grade explained a small fraction of the variance in DLS scores (R2 = 18% in both validation sets). Next, we generated human-interpretable histologic features by clustering embeddings from a deep-learning-based image-similarity model and showed that they explained the majority of the variance (R2 of 73-80%). Furthermore, the clustering-derived feature most strongly associated with high DLS scores was also highly prognostic in isolation. With a distinct visual appearance (poorly differentiated tumor cell clusters adjacent to adipose tissue), this feature was identified by annotators with 87.0-95.5% accuracy. Our approach can be used to explain predictions from a prognostic deep learning model and uncover potentially-novel prognostic features that can be reliably identified by people for future validation studies.
RESUMO
Background: Gleason grading of prostate cancer is an important prognostic factor, but suffers from poor reproducibility, particularly among non-subspecialist pathologists. Although artificial intelligence (A.I.) tools have demonstrated Gleason grading on-par with expert pathologists, it remains an open question whether and to what extent A.I. grading translates to better prognostication. Methods: In this study, we developed a system to predict prostate cancer-specific mortality via A.I.-based Gleason grading and subsequently evaluated its ability to risk-stratify patients on an independent retrospective cohort of 2807 prostatectomy cases from a single European center with 5-25 years of follow-up (median: 13, interquartile range 9-17). Results: Here, we show that the A.I.'s risk scores produced a C-index of 0.84 (95% CI 0.80-0.87) for prostate cancer-specific mortality. Upon discretizing these risk scores into risk groups analogous to pathologist Grade Groups (GG), the A.I. has a C-index of 0.82 (95% CI 0.78-0.85). On the subset of cases with a GG provided in the original pathology report (n = 1517), the A.I.'s C-indices are 0.87 and 0.85 for continuous and discrete grading, respectively, compared to 0.79 (95% CI 0.71-0.86) for GG obtained from the reports. These represent improvements of 0.08 (95% CI 0.01-0.15) and 0.07 (95% CI 0.00-0.14), respectively. Conclusions: Our results suggest that A.I.-based Gleason grading can lead to effective risk stratification, and warrants further evaluation for improving disease management.
RESUMO
Digital pathology is on the verge of becoming a mainstream option for routine diagnostics. Faster whole slide image scanning has paved the way for this development, but implementation on a large scale is challenging on technical, logistical, and financial levels. Comparative studies have published reassuring data on safety and feasibility, but implementation experiences highlight the need for training and the knowledge of pitfalls. Up to half of the pathologists are reluctant to sign out reports on only digital slides and are concerned about reporting without the tool that has represented their profession since its beginning. Guidelines by international pathology organizations aim to safeguard histology in the digital realm, from image acquisition over the setup of work-stations to long-term image archiving, but must be considered a starting point only. Cost-efficiency analyses and occupational health issues need to be addressed comprehensively. Image analysis is blended into the traditional work-flow, and the approval of artificial intelligence for routine diagnostics starts to challenge human evaluation as the gold standard. Here we discuss experiences from past digital pathology implementations, future possibilities through the addition of artificial intelligence, technical and occupational health challenges, and possible changes to the pathologist's profession.