Selver M. A., Barış M. M., Yurt A., Dicle O.
ECR 2025, Vienna, Avusturya, 26 Şubat - 02 Mart 2025, ss.19136-19142, (Tam Metin Bildiri)
-
Yayın Türü:
Bildiri / Tam Metin Bildiri
-
Basıldığı Şehir:
Vienna
-
Basıldığı Ülke:
Avusturya
-
Sayfa Sayıları:
ss.19136-19142
-
Dokuz Eylül Üniversitesi Adresli:
Evet
Özet
Purpose
AI and radiomics-based analyses tend to improve global health equity. Thus, reliability is a key factor in the sustainability of their services. Computed Tomography Texture Analysis (CTTA) provides effective results in diagnosis, treatment, and follow-up for various clinical conditions. Unfortunately, it is also shown that CTTA has low repeatability, and reproducibility, and is highly dependent on modality, acquisition, and reconstruction parameters.For instance, The Credence Cartridge Radiomics (CCR) phantom is a specially designed tool used to evaluate the robustness and variability of radiomics features in computed...
Methods and materials
The phantoms and feature extraction pipeline is illustrated in [Fig 1].Five different texture phantoms, which were 3D-printed using the Raise3D Pro3 Plus 3D printer are analyzed.[Fig 2] The printing material was PLA for all phantoms, each of which had a 4 cm × 4 cm cross-sectional area. Three phantoms were designed as line phantoms, referred to as vertical, horizontal, and diagonal phantoms. One phantom was square-shaped, and the fifth phantom was star-shaped having eight solid slices. A sixth phantom is also designed to have a...
Results
In most of the violin plots, data distributions of the GLCM side propagate widely in terms of spatial domain analysis, whereas the DTCWT side has a more concentrated area for spatial-frequency domain analysis. For diagonal, horizontal, and square textures in the violin plots, the 1st, 2nd, and 3rd quartiles are separable, while honeycomb, sinusoidal, and star textures have close quartiles. In our experimental results, the median value of each DTCWT-based feature type across all textures is close to 1 (as the normalized value) in most...
Conclusion
The repeatability, and reproducibility of SFBT CTTA are more robust to CT acquisition parameters and variations in data characteristics compared to image domain (i.e. GLCM) texture features. The SFBT and image domain texture features may have similar robustness only when the texture is rotation invariant or symmetric with respect to its origin. Since textures in real medical images mostly do not have such symmetry and rotation invariance properties, SFBT CTTA features should be preferred for radiomics analyses.