Radiomics, the process of extracting vast quantities of quantitative data from medical images, holds great potential for revolutionising modern healthcare. It offers applications ranging from disease diagnosis to patient prognosis and stratification. Despite its promise, the clinical utility of radiomics has been stifled by one critical issue: reproducibility. Variations in imaging protocols, scanner models and reconstruction techniques often lead to inconsistencies in the extracted radiomics features, limiting their reliability. Deep learning-based harmonisation has emerged as an innovative solution, aiming to standardise image quality and feature consistency across diverse datasets. A recent review published in Bioengineering explores the use of deep learning models to improve radiomics reproducibility in abdominal computed tomography (CT) scans.
The Challenge of Radiomics Reproducibility
Radiomics relies on extracting consistent and meaningful features from medical images, yet multiple factors undermine its reproducibility. Differences in imaging parameters, such as slice thickness, reconstruction algorithms, and scanner types, introduce variability in feature extraction. This variability complicates multi-centre studies, making comparing findings or applying insights universally difficult. Furthermore, changes in preprocessing steps exacerbate the issue, leading to unstable models that are less likely to perform reliably in clinical practice.
Efforts to address these challenges have largely focused on standardisation protocols. Organisations such as the Food and Drug Administration (FDA) and the European Association of Nuclear Medicine (EANM) have established guidelines for imaging consistency. However, these guidelines are often project-specific, lack adaptability for retrospective studies and fail to account for changing hardware and software. Consequently, there is a growing demand for dynamic solutions that can adapt to heterogeneous imaging datasets while preserving the integrity of radiomics features.
Deep learning represents a promising approach to mitigating these challenges. Generative adversarial networks (GANs) have demonstrated the ability to harmonise imaging data by correcting inconsistencies and standardising outputs. In controlled settings, these networks have successfully enhanced the reproducibility of synthetic datasets. However, the transition from phantom studies to real-world clinical settings requires extensive validation using patient-derived imaging data. This study takes a step forward by evaluating the application of deep learning-based harmonisation in real patient CT scans.
Deep Learning for Harmonisation
The study employed a deep-learning model designed to harmonise CT images across various acquisition protocols. The model utilised hierarchical feature synthesis and sequential spatial-channel attention mechanisms to identify and correct inconsistencies. Internal validation demonstrated a remarkable improvement in radiomics reproducibility, with the percentage of features meeting the reliability threshold (intraclass correlation coefficient ≥0.85) rising from 18% to 65%. External validation results mirrored these findings, confirming the model's robustness across datasets.
One of the most significant improvements was observed in the reproducibility of features from complex anatomical regions, such as vessels and organs like the kidneys and spleen. Vessel features, for instance, saw reproducibility rates rise from 14% to 69% following harmonisation. Other regions, such as the liver parenchyma and spleen, also experienced notable enhancements. While air regions exhibited a minor decrease in reproducibility, this had negligible impact on the overall feature set. These results underline the model's ability to address variability in regions that are often challenging to standardise.
Unlike traditional approaches, this model excluded redundant features—those that fail to capture meaningful variations across protocols. By focusing solely on clinically relevant features, the method optimises the analysis and improves the efficiency of radiomics workflows. This innovation has broad implications for the scalability of radiomics in both research and clinical applications, particularly in multi-centre studies where imaging protocols vary significantly.
Another key advantage is the adaptability of deep learning-based harmonisation. Unlike standardisation guidelines, which often require rigid adherence during the image acquisition phase, this approach allows for post hoc harmonisation. This is especially valuable for retrospective datasets, which frequently lack the uniformity required for robust radiomics analysis. The flexibility of deep learning models thus provides a powerful tool for overcoming the limitations of traditional standardisation methods.
Implications for Clinical and Research Applications
The implications of this study extend beyond technical enhancements, offering meaningful advancements for clinical and research applications. By improving reproducibility, the model enables radiomics to be used more confidently in large-scale studies and clinical trials. This reliability is essential for generating predictive models, identifying biomarkers and making data-driven clinical decisions.
In clinical practice, the enhanced reproducibility of radiomics features can facilitate more accurate diagnoses and personalised treatment plans. For instance, standardised radiomics features could be used to monitor disease progression or predict treatment responses with greater confidence. Moreover, the ability to harmonise datasets retrospectively ensures that older studies can contribute to ongoing research without compromising data integrity.
Despite these promising developments, the study also highlights several limitations that must be addressed in future research. The external validation was conducted using a single type of CT scanner and a limited range of reconstruction protocols, which may limit the generalisability of the findings. Expanding validation efforts to include diverse scanners, protocols and pathological conditions will be essential to confirm the model's applicability across a wider range of clinical scenarios. Furthermore, the study did not assess interobserver variability, an important consideration given the impact of manual segmentation on radiomics features.
Another limitation lies in the absence of direct evaluations involving pathological lesions. Real-world clinical applications often involve analysing lesions, which may present additional challenges not addressed in this study. As such, future work should focus on testing the model in scenarios involving true pathological variability and observer-dependent processes.
Deep learning-based harmonisation represents a pivotal advancement in improving the reproducibility of radiomics features in abdominal CT imaging. This innovative approach ensures greater consistency and reliability in radiomics analysis by addressing variability in imaging protocols and preprocessing steps. While challenges remain, including the need for broader validation and assessments of interobserver variability, the results of this study underscore the transformative potential of deep learning in radiomics. This technology is expected to unlock the full capabilities of radiomics, unlocking more personalised, data-driven approaches to healthcare.
Source: Bioengineering
Image Credit: iStock
