To Distinguish AC and SCC In Cervical Cancer Using Machine Learning Models Based on Radiological Features of 18F-FDG PET / CT
Abstract
Purpose
To determine the diagnostic performance of a machine learning model based on radiographic features of fluorine-18-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET) / computed tomography (CT) in distinguishing cervical adenocarcinoma (AC) from squamous cell carcinoma (SCC).
Methods
18F-FDG PET / CT data from 227 patients diagnosed with locally advanced cervical cancer were retrospectively collected from two centers. Inclusion criteria included: pathologic diagnosis of cervical cancer and compliance with the 2018 International Union of Obstetrics and Gynecology (FIGO) IB-IVA, 18F-FDG PET / CT, and complete retrieval of clinical data. PET images were subjected for attenuation correction, reconstruction and multifaceted fusion using MIM Maestro software and fused with non-contrast-enhanced low-dose CT images. Bladder region were manually excluded on the original images and parameters such as metabolically active tumor volume (MTV), mean normalized uptake value (SUVmean), total lesion glycolysis (TLG) and SUVmax were calculated.
Results
In the training cohort, the optical gradient elevator (lightGBM) model based on PET radiology features showed the best discriminatory performance (AUC = 0.955), followed by the AUC of the combined model (AUC = 0.968). In the in-house validation cohort, the AUC of the PET radiographic model was 0.851 and the AUC of the binding model was 0.842. In the external validation cohort, the AUC of the PET radiographic model was 0.730. The DeLong test showed no significant difference in the AUC between the combined model and PET radiology models in the training and in-house validation cohorts, but both were significantly better than the CT radiology model.
Conclusion
The optical gradient hoist model based on 18F-FDG PET radiographic features able to effectively distinguish pathological subtypes of locally advanced cervical cancer may facilitate the daily decision-making process of clinicians. This study demonstrates the potential of radiological features in distinguishing subtypes of tumor pathology and provides direction for future studies.