Abstracts

RETURN TO ABSTRACT LISTING


E1398. Texture Analysis of ADC Maps for Differentiating Between Fat-Poor AML and Non-Clear Cell RCC: Model Development and External Validation
Authors
  1. Yuki Arita; Keio University School of Medicine
  2. Shigeo Okuda; Keio University School of Medicine
  3. Hirotaka Akita; Keio University School of Medicine
  4. Soichiro Yoshida; Tokyo Medical and Dental University Graduate School
  5. Ryo Ueda; Keio University School of Medicine
  6. Ryota Ishii; Keio University School of Medicine
  7. Masahiro Jinzaki; Keio University School of Medicine
Objective:
Since approximately 20% of the solid small renal masses < 4 cm are benign, warranting conservative management, preoperative imaging should aim to differentiate benign from malignant tumors [1]. Magnetic resonance imaging (MRI) techniques have also been explored for this purpose, including the use of advanced techniques, such as diffusion-weighted imaging (DWI) [2,3]. Although these methods are promising, they remain suboptimal and have not found clinical application. For example, calculated apparent diffusion coefficient (ADC) values may vary for different magnetic resonance (MR) devices or settings. This study aimed to determine the feasibility of texture analysis (TA) of ADC maps for differentiating between fat-poor angiomyolipomas (fpAMLs) and non-clear cell RCCs (non-ccRCCs). Furthermore, the diagnostic performance of the random forest (RF) model generated using selective texture features (TFs) obtained from the development cohort was validated in an external validation cohort.

Materials and Methods:
This study included two cohorts (from different institutions) with renal masses (with a final pathological diagnosis). Patients underwent MRI, including 1.5-T DWI, before surgical intervention. MR devices from different manufacturers were used. Sixty-seven cases (46 fpAMLs and 21 non-ccRCCs) were included in model development and 39 (24 fpAMLs and 15 non-ccRCCs) cases were included for validation. Forty-five TFs, including higher-order TFs, were extracted using software with volumes of interest on ADC maps generated from DWI using b-values of 0 and 1000 s/mm2. Receiver operating characteristic analysis was performed to compare the diagnostic performance of the RF model and ADC values for discriminating fpAMLs from non-ccRCCs.

Results:
The median tumor size (mm) was significantly larger in the development cohort than that in the validation cohort (31 versus 22; P < 0.001). RF analysis revealed that grey-level zone length matrix_long-zone high gray-level emphasis (GLZLM_LZHGE), which indicates the extent of signal homogeneity, was the dominant TF for differentiating between fpAMLs and non-ccRCCs. Internal cross-validation revealed that the area under the curve (AUC) for distinguishing between fpAMLs and non-ccRCCs was significantly larger for the RF model than that for ADC values (0.83 versus 0.61; P < 0.001). External validation showed that the AUC for distinguishing between fpAMLs and non-ccRCCs was significantly larger for the RF model than that for ADC values (0.82 versus 0.52; P < 0.001). No significant difference was found between the AUCs for the intracross and external validation cohorts in the RF model (0.83 versus 0.82; P = 0.25).

Conclusion:
TA of the ADC maps demonstrated a significantly higher diagnostic performance than that of ADC values for differentiating between fpAMLs and non-ccRCCs. The RF model provided similar diagnostic performance for distinguishing between fpAMLs and non-ccRCCs in both intracross and external validation cohorts, despite the difference between patient cohorts and MR scanner manufacturers.