- •
Artificial intelligence (AI) techniques are being increasingly explored in medical imaging. Innovations in machine learning (ML) and deep learning (DL) have helped unlock the potentials of AI for successful applications.
- •
Patient health information, including demographic information, electronic medical record notes, diagnostic imaging at different time-points, and radiologist reports, along with radiomic features can be used as input to AI techniques for detection, classification, and outcome
AI-Based Detection, Classification and Prediction/Prognosis in Medical Imaging: Towards Radiophenomics
Section snippets
Key points
Advanced image quantification; the field of radiomics
Medical images contain a significant minable and potentially valuable quantitative information beyond what is nowadays captured in routine clinical evaluations, motivating the field of radiomics.28,29 An array of AI techniques in the field of medical imaging have emerged in the past decade to derive imaging biomarkers based on this information, using explicit (ie, handcrafted/engineered) radiomics features or deep radiomics features (ie, derived via deep neural networks (NNs)). These techniques
Radiomics signatures
Radiomics analyses can augment visual assessments made by radiologists.44 AI techniques can perform quantitative high-throughput image phenotyping (extracting numerous image-based features) and identifying important discriminative features that individually or in combination form an effective radiomics signature for a given task. The field of radiomics has been introduced and elaborated in an accompanying article by Orlhac and colleagues.29 The challenging task of detecting suspicious regions
Clinical utility of artificial intelligence-based detection in positron emission tomography/computed tomography: current status
Detection of organs or lesions as regions of interests (ROIs) or volumes of interests (VOIs) is an important step toward the classification of the regions.69 Besides, most existing segmentation approaches have an embedded detection step; we discuss segmentation methods in another work [see Yousefirizi and colleagues’ article, “Towards High-throughput AI-based Segmentation in Oncological PET Imaging,” in this issue]. End-to-end detection systems developed by AI techniques can effectively remove
Clinical utility of artificial intelligence-based classification in positron emission tomography/computed tomography: current status
Classification is considered the most popular area in which CNNs have been used; for example, AlexNet, ResNet, DenseNet, VGG network, and others.79 Similarly, in medical imaging, AI techniques have been widely used for the extraction of feature toward: (i) classification of suspicious lesions and tumor subtypes, as well as (ii) prediction/prognostication tasks, stratifying/classifying patients into risk groups.111,112 We review both frontiers next.
Temporal changes in radiomics features: dynamic-radiomics and delta-radiomics
Radiomics signatures of medical images are usually derived from static images. It has been shown that temporal changes for example in tracer uptake of PET scans can also reveal new aspects of tumor biology.142, 143, 144, 145 Analysis of features extracted from the temporal analysis of dynamic PET scans can be referred to as “dynamic-radiomics” (micro-scale temporal changes). By contrast, “delta-radiomics” refer to the analysis of features derived from the comparison of a study with prior images
Summary
Considering the strength of AI techniques in performing effective image phenotyping including robust identification of patterns beyond visual assessments, there is significant potential for use in accelerating and automating detection and classification tasks in medical imaging and as an example oncologic PET imaging, as reviewed in the present work. AI techniques can also link radiomics signatures and biological properties extracted by radiologists and nuclear medicine physicians in the
Clinics care points
Stratification of cancer into reliably distinct risk subgroups enables personalization of treatment. Detection of organs or lesions as regions of interests (ROIs) or volumes of interests (VOIs) is an important step towards classification. Classification tasks for identification and characterization of lesions have significant potential in clinical workflows. Classification in the context of prediction/prognosis also enables risk stratification. Radiomics mines for potentially significant
Acknowledgements
This project was in part supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant RGPIN-2019-06467, and the Canadian Institutes of Health Research (CIHR) Project Grant PJT-173231.
Disclosure
Authors do not have anything to disclose regarding conflict of interest with respect to this article.
References (194)
A new framework for understanding vision from the perspective of the primary visual cortex
Curr Opin Neurobiol
(2019)- et al.
Untangling invariant object recognition
Trends Cognitive Sciences
(2007) - et al.
How does the brain solve visual object recognition?
Neuron
(2012) - et al.
Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multireader study
Lancet Digital Health
(2020) - et al.
A network-based “phenomics” approach for discovering patient subtypes from high-throughput cardiac imaging data
JACC: Cardiovasc Imaging
(2020) - et al.
Genetics and phenomics of Pendred syndrome
Mol Cell Endocrinol
(2010) - et al.
Artificial intelligence and machine learning in nuclear medicine: future Perspectives
Semin Nucl Med
(2021) - et al.
Radiomics in PET imaging: a practical guide for newcomers
PET Clin
(2021) - et al.
Exploring feature-based approaches in PET images for predicting cancer treatment outcomes
Pattern recognition
(2009) - et al.
FDG-PET-based prognostic nomograms for locally advanced cervical cancer
Gynecol Oncol
(2012)