Keynote: Radiomics and artificial intelligence are revolutionising oncologic imaging by enabling non-invasive, data-driven insights into tumour biology, but standardisation, harmonisation, and large-scale collaboration remain essential for clinical translation.
Keywords: Radiomics, artificial intelligence, deep learning, oncologic imaging, feature extraction, clinical translation
Radiomics and artificial intelligence (AI) offer considerable promise for enhancing oncologic imaging by extracting and analysing thousands of features embedded in conventional imaging data. These features—quantifiable representations of tissue biology, structure, and function—can serve as non-invasive imaging biomarkers that support cancer detection, characterisation, prognosis, and therapy response prediction.
The radiomics workflow involves image acquisition, segmentation, feature extraction, model training, and validation. Standardisation at every stage is essential to reduce non-biological variability, such as scanner differences or image reconstruction settings, which can undermine model reproducibility and generalisability. Harmonisation techniques—including image pre-processing and feature-level adjustments, such as ComBat—have been developed to mitigate batch effects, particularly when working with multicentre datasets.
AI plays a key role throughout the radiomics pipeline. Machine learning (ML) models can classify and predict outcomes based on hand-crafted features, while deep learning (DL) offers an alternative that bypasses manual feature engineering through end-to-end feature learning. DL models, particularly convolutional neural networks (CNNs), are being explored for automated segmentation, lesion detection, and image classification.
Radiomics has shown potential across multiple clinical applications: improving cancer screening accuracy, characterising lesion histology, assessing the tumour microenvironment, and predicting treatment response and recurrence. For example, radiomic signatures have been developed to predict immune-related therapy response, tumour hypoxia, and long-term survival in patients with glioblastoma or lung cancer.
However, several barriers hinder clinical translation. Variability in imaging protocols, segmentation methods, and feature extraction workflows remains a major concern. Moreover, the interpretability of complex models—particularly DL-based ones—limits clinical confidence. Small, retrospective datasets dominate the field, and the lack of multicentre, prospective studies hampers the development of robust, generalisable models.
Efforts to address these issues include standardising imaging protocols, improving automated segmentation tools, and fostering multi-institutional collaboration. Federated learning, which allows decentralised model training while preserving data privacy, is emerging as a viable approach to scale AI across healthcare systems. Harmonisation and reproducibility remain critical for the acceptance of radiomics in clinical trials and real-world practice.
In conclusion, radiomics and AI hold transformative potential for personalised cancer care. Their successful clinical adoption will require rigorous standardisation, collaboration, and continued innovation in data processing and model validation.
Disclaimer
This summary is intended for informational and educational purposes only. It is based on the original open-access article by Majumder et al., published in BJR|Open (January 2024), and does not substitute for professional medical advice, diagnosis, or treatment. The content does not reflect the official views of the authors, journal, or affiliated institutions. Readers are encouraged to consult the original publication for full context and to critically evaluate the information in light of evolving research and clinical guidelines.