{"title":"THE ROLE OF ARTIFICIAL INTELLIGENCE IN IMAGING READINGS","authors":"I. Buvat","doi":"10.1002/hon.70093_9","DOIUrl":null,"url":null,"abstract":"<p>Artificial intelligence (AI) is gaining ground in medical imaging thanks to the increasing availability of open datasets and shared deep learning models. In the context of imaging readings, it can mainly serve two purposes. The first is to automate the detection of abnormalities and the extraction of quantitative features from the images. The second is to predict the future of the patient based on image content possibly supplemented by clinical, pathological and/or biological information.</p><p>In this talk, we will show that AI can already be used to automate a number of tedious tasks often prone to intra- and inter-reader variability, such as lesion detection and segmentation from whole-body [18F]-FDG PET/CT images. This enables automated calculation of prognostic biomarkers from these images, such as the total metabolically active tumor volume, and exploration of the prognostic or predictive values of numerous candidate radiomic biomarkers. We will also discuss the variability between different AI algorithms, requiring the establishment of benchmarks to determine the performance of each AI algorithm and its compliance with interpretation rules agreed by medical experts.</p><p>In a second part, we will present the challenging task of predicting treatment response or patient outcome based on image readings. We'll explain how AI can help make the most of image content. The differences between using end-to-end deep learning and using radiomic features associated with machine learning will be explained, highlighting the advantages and limitations of each approach for prediction tasks. In addition to medical images, the inclusion of non-imaging data in prognostic and predictive models may be necessary to improve performance. We will illustrate how this can be achieved. The challenges associated with using AI for inference will be described based on examples from the literature and our own experience.</p><p><b>Keywords:</b> diagnostic and prognostic biomarkers; PET-CT; risk models</p><p>No potential sources of conflict of interest.</p>","PeriodicalId":12882,"journal":{"name":"Hematological Oncology","volume":"43 S3","pages":""},"PeriodicalIF":3.3000,"publicationDate":"2025-06-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/hon.70093_9","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Hematological Oncology","FirstCategoryId":"3","ListUrlMain":"https://onlinelibrary.wiley.com/doi/10.1002/hon.70093_9","RegionNum":4,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"HEMATOLOGY","Score":null,"Total":0}
引用次数: 0
Abstract
Artificial intelligence (AI) is gaining ground in medical imaging thanks to the increasing availability of open datasets and shared deep learning models. In the context of imaging readings, it can mainly serve two purposes. The first is to automate the detection of abnormalities and the extraction of quantitative features from the images. The second is to predict the future of the patient based on image content possibly supplemented by clinical, pathological and/or biological information.
In this talk, we will show that AI can already be used to automate a number of tedious tasks often prone to intra- and inter-reader variability, such as lesion detection and segmentation from whole-body [18F]-FDG PET/CT images. This enables automated calculation of prognostic biomarkers from these images, such as the total metabolically active tumor volume, and exploration of the prognostic or predictive values of numerous candidate radiomic biomarkers. We will also discuss the variability between different AI algorithms, requiring the establishment of benchmarks to determine the performance of each AI algorithm and its compliance with interpretation rules agreed by medical experts.
In a second part, we will present the challenging task of predicting treatment response or patient outcome based on image readings. We'll explain how AI can help make the most of image content. The differences between using end-to-end deep learning and using radiomic features associated with machine learning will be explained, highlighting the advantages and limitations of each approach for prediction tasks. In addition to medical images, the inclusion of non-imaging data in prognostic and predictive models may be necessary to improve performance. We will illustrate how this can be achieved. The challenges associated with using AI for inference will be described based on examples from the literature and our own experience.
Keywords: diagnostic and prognostic biomarkers; PET-CT; risk models
期刊介绍：
Hematological Oncology considers for publication articles dealing with experimental and clinical aspects of neoplastic diseases of the hemopoietic and lymphoid systems and relevant related matters. Translational studies applying basic science to clinical issues are particularly welcomed. Manuscripts dealing with the following areas are encouraged:
-Clinical practice and management of hematological neoplasia, including: acute and chronic leukemias, malignant lymphomas, myeloproliferative disorders
-Diagnostic investigations, including imaging and laboratory assays
-Epidemiology, pathology and pathobiology of hematological neoplasia of hematological diseases
-Therapeutic issues including Phase 1, 2 or 3 trials as well as allogeneic and autologous stem cell transplantation studies
-Aspects of the cell biology, molecular biology, molecular genetics and cytogenetics of normal or diseased hematopoeisis and lymphopoiesis, including stem cells and cytokines and other regulatory systems.
Concise, topical review material is welcomed, especially if it makes new concepts and ideas accessible to a wider community. Proposals for review material may be discussed with the Editor-in-Chief. Collections of case material and case reports will be considered only if they have broader scientific or clinical relevance.
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号
