MATCHING CT IMAGE CHARACTERISTICS TO IMPROVE RADIOMIC REPRODUCIBILITY: A TRANSLATIONAL CALIBRATION PIPELINE USING 3D-PRINTED PHANTOMS:
Radiological Society of North America (RSNA), Illinois, Amerika Birleşik Devletleri, 29 Kasım - 03 Aralık 2025, ss.170, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: Illinois
- Basıldığı Ülke: Amerika Birleşik Devletleri
- Sayfa Sayıları: ss.170
- Dokuz Eylül Üniversitesi Adresli: Evet
Özet
urpose: To enhance the clinical integration of radiomics by improving the reproducibility of CT-derived features through real-time correction of acquisition variability using a 3Dprinted phantom-based calibration. *Methods and Materials: Twenty cylindrical inserts (diameter and height: 4cm) spanning -400 to +400 HU were scanned using a single CT (Philips Brilliance 64-slice) and imaging protocol (head/neck). Central ROls (radius = 20 pixels) were taken from each. Mean and standard deviation values were used as reference points for intensity and noise calibration. Alongside these, ten 3D-printed texture phantoms (4 x 4 x 4 cm) with diverse texture patterns were included in each scan. Over six months, 24 acquisitions were performed to simulate scanner drift. A piecewise linear intensity stretching method was applied to match the mean of reference cylinders across scans, correcting for data drift. Standard deviation differences are used to determine level of noise removal or insertion. IBSI-compliant radiomic features were extracted from each texture before and after calibration. Feature variability was quantified using Normalized Standard Deviation (NSD), and image similarity is assesed by Structural Similarity Index Measure (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS) across timepoints. *Results: Prior to calibration, high variability was observed across repeated scans, particularly in frequency- and low-level texture features (e.g., DCT, GLCM). Following calibration, median NSD across all features and textures dropped from 0.28 to 0.09, representing more than 70% improvement in feature reproducibility. Moreover, without calibration, the observed SSIM and LPIPS were 0.87 and 0.17 in average, respectively, which is increased to 0.95 for SSIM, while LPIPS decreased to 0.11. These improvements show that the proposed calibration pipeline effectively minimizes inter-scan radiomics variability while preserving texture-specific signals. *Conclusions: This study introduces a practical, scanner-agnostic calibration framework using 3D-printed phantoms to align CT image characteristics across acquisitions. By significantly enhancing radiomic feature reproducibility, the method addresses a key barrier to clinical and regulatory adoption of quantitative imaging biomarkers. *Clinical Relevance/Application: Radiomic features are increasingly used in Al models, prognostic algorithms, and clinical trials. Yet their clinical utility is undermined by poor reproducibility across scanners and timepoints. The proposed method is the first to combine a custom CT phantom with real-time data drift detection and mitigation algorithms, enabling scanner-specific, patient-centric calibration of CT images before quantitative analysis.