The [[sup.18]F]F-FDG PET/CT Radiomics Classifier of Histologic Subtypes and Anatomical Disease Origins across Various Malignancies: A Proof-of-Principle Study

There is a lack of ample knowledge about the origin of anatomical disease- and histology-specific radiomic variations across different malignancies. Thus, the aim of our proof-of-principle study was to investigate whether radiomics features have the ability to discriminate between histological subtypes and to predict the anatomical disease origin of different tumour entities. Also, we tried to synchronously predict histology and anatomical disease origin using baseline [[sup.18]F]F-FDG PET/CT. Based on our findings, our proof-of-principle study may demonstrate a potentially high degree of diagnostic accuracy to predict histology and disease origin across different tumour entities, using the standard of care combined [[sup.18]F]F-FDG PET/CT-derived radiomics features. This may help with the development of a new paradigm regarding the potential of also thinking about tumours from their molecular basis, and not only on the basis of their anatomical origin perse. We aimed to investigate whether [[sup.18]F]F-FDG-PET/CT-derived radiomics can classify histologic subtypes and determine the anatomical origin of various malignancies. In this IRB-approved retrospective study, 391 patients (age = 66.7 ± 11.2) with pulmonary (n = 142), gastroesophageal (n = 128) and head and neck (n = 121) malignancies were included. Image segmentation and feature extraction were performed semi-automatically. Two models (all possible subset regression [APS] and recursive partitioning) were employed to predict histology (squamous cell carcinoma [SCC; n = 219] vs. adenocarcinoma [AC; n = 172]), the anatomical origin, and histology plus anatomical origin. The recursive partitioning algorithm outperformed APS to determine histology (sensitivity 0.90 vs. 0.73; specificity 0.77 vs. 0.65). The recursive partitioning algorithm also revealed good predictive ability regarding anatomical origin. Particularly, pulmonary malignancies were identified with high accuracy (sensitivity 0.93; specificity 0.98). Finally, a model for the synchronous prediction of histology and anatomical disease origin resulted in high accuracy in determining gastroesophageal AC (sensitivity 0.88; specificity 0.92), pulmonary AC (sensitivity 0.89; specificity 0.88) and head and neck SCC (sensitivity 0.91; specificity 0.92). Adding PET-features was associated with marginal incremental value for both the prediction of histology and origin in the APS model. Overall, our study demonstrated a good predictive ability to determin