Extended abstract: radiomics and artificial intelligence in thymic tumors

Radiomics has arisen as a burgeoning field of translational research predicated on the acquisition of high-dimensional data from radiological images, with the objective of discerning distinctive imaging patterns that may not be readily perceptible to the human observer.

The utility of radiomics extends across several critical applications. It can serve as a diagnostic tool capable of ascertaining the overall normalcy or abnormality of an imaging study under scrutiny, rendering it an invaluable asset for screening purposes. Furthermore, radiomics can pinpoint the precise location of pathological conditions within a study and exclude the presence of specific diseases. Additionally, among other applications, it can contribute to the enhancement of image quality, such as through contrast optimization.

Radiomics features are derived via a meticulous and structured process. The initial step involves segmenting the area of interest, typically a lesion, by delineating its boundaries on each image slice. This segmentation process can be accomplished either manually, although this method is time-intensive, or automatically or semi-automatically through algorithmic approaches, often necessitating subsequent manual refinement for precision.

While it is feasible to extract radiomics features from individual slices, it is commonplace to derive these features across all image slices, thus yielding a comprehensive dataset representing the entire volume of interest. Various methodologies exist for feature extraction, which can be categorized into statistical, model-based, and transform-based methods, each providing distinct types of variables. Python is the predominant software platform employed for radiomics feature extraction and subsequent analysis.

First-order statistical radiomics features encapsulate the distribution of grayscale values in a histogram, encompassing metrics like entropy, skewness, and kurtosis. Entropy quantifies the stochasticity or randomness within the pixel intensity distribution, while skewness characterizes its asymmetry, and kurtosis gauges the shape of the histogram.

Second-order statistical radiomics features establish mathematical relationships between neighboring pixels and are expressed in terms of neighborhood matrices, including entities like gray level cooccurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone matrix (GLSZM).

Third-order radiomics features, such as harmonization and wavelet transformations, contribute to capturing texture and spatial information, while fourth-order features, associated with convolutional neural networks (CNNs), enable the extraction of high-level representations for more advanced and nuanced analysis of medical images.

Upon the extraction of radiomics features, a plethora of variables is generated. Feature selection is a pivotal step in the process, involving the curation of the most pertinent attributes to enhance analytical precision and the development of efficient predictive models. This task may encompass traditional statistical methodologies, chiefly logistic regression models, as well as advanced artificial intelligence (AI) analysis.

Machine learning, a subset of AI, demonstrates superior proficiency in handling extensive datasets and exhibits improved diagnostic performance compared to traditional statistical approaches. For AI analysis, the dataset is divided into training and validation subsets, which encompass 60–70% of the data, and a distinct test subset with the remaining 30–40%. Machine learning algorithms are employed during the training phase to construct models tailored to the study’s objectives, and these models are subsequently evaluated for performance. Adjustments to hyperparameters are made if the obtained area under the curve (AUC) falls short of expectations. The final assessment of model performance occurs when tested against an independent, previously unseen dataset, which provides a reliable estimate of the diagnostic accuracy of the model.

In the context of thymic tumors, radiomics with AI analysis hold substantial promise across various domains. Our recent published study, conducted by the chest radiology team at Memorial Sloan Kettering Cancer Center, aimed to develop radiomics and AI models for distinguishing between benign and malignant prevascular mediastinal lesions, as well as differentiating between thymomas and thymic carcinomas in preoperative computed tomography (CT) studies (1). When relying solely on conventional CT features, the diagnostic performance of these models proved suboptimal. However, the inclusion of radiomics features in the models significantly enhanced their AUC, with the highest performance achieved when both conventional and radiomics features were combined. Specifically, this integration resulted in an AUC value of 0.72 for distinguishing between benign and malignant anterior mediastinal lesions, and an AUC of 0.81 for differentiation between thymomas and thymic carcinomas.

In a previously published study, we examined the utility of radiomics in predicting the resectability and stage of thymic tumors (2). Our findings indicated a robust diagnostic performance for resectability prediction, with an AUC of 0.80, and a moderate discriminative ability for stage prediction, with an AUC of 0.71. These results align with a recent meta-analysis that included our study (3). In this meta-analysis, the evaluation of combined diagnostic performance for stage prediction of thymic tumors yielded a pooled AUC of 0.83 (for distinguishing between early and advanced disease) and a combined AUC of 0.86 for predicting histologic subtypes (for discriminating between low and high-risk tumors).

In conclusion, radiomics, coupled with AI analysis, represents a reliable and potent tool for predicting risk stratification, stage, and resectability of thymic tumors, offering the potential to enhance the preoperative evaluation of these conditions and advancing the goal of personalized medicine.

Acknowledgments

Funding: None.

Footnote

Provenance and Peer Review: This article was commissioned by the Guest Editors (Malgorzata Szolkowska, Chul Kim, Mohammad Ashraghi, and Claudio Silva) for “The Series Dedicated to the 13th International Thymic Malignancy Interest Group Annual Meeting (ITMIG 2023)” published in Mediastinum. The article has undergone external peer review.

Peer Review File: Available at https://med.amegroups.com/article/view/10.21037/med-23-50/prf

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://med.amegroups.com/article/view/10.21037/med-23-50/coif). “The Series Dedicated to the 13th International Thymic Malignancy Interest Group Annual Meeting (ITMIG 2023)” was commissioned by the editorial office without any funding or sponsorship. The author has no other conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Mayoral M, Pagano AM, Araujo-Filho JAB, et al. Conventional and radiomic features to predict pathology in the preoperative assessment of anterior mediastinal masses. Lung Cancer 2023;178:206-12. [Crossref] [PubMed]
2. Araujo-Filho JAB, Mayoral M, Zheng J, et al. CT Radiomic Features for Predicting Resectability and TNM Staging in Thymic Epithelial Tumors. Ann Thorac Surg 2022;113:957-65. [Crossref] [PubMed]
3. Lu XF, Zhu TY. Diagnostic performance of radiomics model for preoperative risk categorization in thymic epithelial tumors: a systematic review and meta-analysis. BMC Med Imaging 2023;23:115. [Crossref] [PubMed]