Applications of Radiomics for Next-Generation Oncologic Management – Current Trend, Challenges and Future Prospects

The rapid advancements of AI and computational technologies have tremendously driven the soaring demand for personalized medicine, particularly in the field of oncology. The contemporary “-Omics” era encompasses genomics, proteomics, metabolomics, etc., which have made dominant and immense impacts on personalized healthcare delivery for the past couple decades, while an aborning concept of radiomics has been introduced since a decade ago, which involves a high-throughput extraction of quantitative, standardized, and voxel-based imaging features, including but not limited to tissue morphology, first-order statistics and spatial-related heterogeneity/texture, from radiographic images. Neoplasms are intrinsically heterogenous and contain multiple sub-clusters of cancer subpopulations; therefore, imaging textures representing clusters of adjacent imaging voxels can be grouped together to derive metrics for tumor subpopulations or habitats that are more representative and objective in reflecting intra-tumoral heterogeneity.

Over the past decade, mounting evidence has demonstrated the superiority of radiomics over conventional qualitative radiologic, histopathologic and clinical attributes from virtual biopsy, cancer staging, cancer histological classification, cancer prognostication, disease differentiation (e.g., pseudo-progression vs cancer recurrence), radiation-induced toxicity prediction, identification of patients who may be reluctant to neo-adjuvant chemotherapy and immunotherapy, etc. Furthermore, the term “delta-radiomics” has emerged as to reflect the dynamic temporal changes in radiomic features, that may capture treatment response or cancer progression patterns that would otherwise not be measurable using current practice. The predictive power of delta-radiomics has been demonstrated to outperform radiomic features from static images. Apart from this, a growing amount of research has illustrated certain radiomic features have been highly correlated with existing genomic markers along with expression of various microRNA signature associated with tumor response to treatment perturbations, cancer metastatic spread, and prognosis; integrating both radiomics and genomics have paved the way toward the nascent area of “radio-genomics” within the community. Moreover, there are also growing number of research on sub-regional radiomics, reporting that peri-tumoral radiomics yielded a greater predictive power than tumor-core radiomics in identifying at-risk patients of post-treatment cancer metastases. There are lot more exciting and innovative radiomics research in the current body of literature. Without doubt, radiomics offers immense and tantalizing potential to serve as a supplementary technique to the existing methods, and to revolutionize cancer management toward personalized oncologic care delivery.

Notwithstanding, there exist several caveats of radiomics, which if addressed, will gain further confidence and trust from clinical practitioners towards model bench-to-bedside translation. Key stumbling blocks include the lack of standardized radiomic workflow and clear reporting of study methodologies, radiomic feature reproducibility across image acquisition protocols or scanner vendors, tumor delineation, etc., lack of large-cohort data for effective model development and external cohort for model external validation (partly due to the practical concern of patient privacy protection), the common circumstances of highly imbalanced data in the field of oncology, challenges in maximizing the harvest of complementary predictive features between various -omics data and/or radiomics features from different imaging modalities/sequences; whether or not the model is explainable, whether or not the accuracy of the radiomic predictors can be generalized to tumors in extreme sizes, recurrent lesions, patients with multiple metastasis, etc.

To this end, the research community has been inventing solutions to the above challenges. For instance, several guidelines for standardized radiomic features extraction, checklists for reporting in radiomics study, radiomic quality scores for assessing the study design of radiomics research. Besides, several radiomic feature reproducibility assessment approaches have been developed and advocated recently for safeguarding model generalizability in unseen populations. Also, an emerging strategy called federated learning, which aims to get rid of the concern of patient privacy disclosure during model development, has been reported in recent years. Various data imbalanced adjustment frameworks as well as sophisticated techniques for multi-omics / multi-view fusion have been developed and reported in the literature.

To conclude, radiomics is playing an influential role as part of next-generation oncologic management. Although it is still in its infant stage in history, tremendous and concerted efforts have been constantly made to revolutionize the role of radiomics in personalized oncology. The developmental pathway and potential of genomics can be an analogy to those of radiomics, with a confident hope that radiomics can eventually assist in routine clinical decision-making in oncology, and it is highly anticipated that the synergistic power of radio-genomics will ultimately generate game-changing impacts in the long run. Nevertheless, it is highly imperative to first create clinical awareness of the concepts of radiomics, hence driving further translational research and clinical trials. Like any other inventions, maturity comes with time, experience, and creativity from eminence worldwide. Global solidarity is the key to success!