2923. Precise Correlation of MRI and Surgical Pathology for the Identification of Aggressive Tumor Features in Renal Cell Carcinoma
Authors* Denotes Presenting Author
Madhu Gowda *;
University of Wisconsin School of Medicine and Public Health
E Jason Abel;
University of Wisconsin School of Medicine and Public Health
Ruben Ngnitewe Massa'a;
University of Wisconsin School of Medicine and Public Health
Daniel Shapiro;
University of Wisconsin School of Medicine and Public Health
Andrew Wentland;
University of Wisconsin School of Medicine and Public Health
Rong Hu;
University of Wisconsin School of Medicine and Public Health
MEghan Lubner;
University of Wisconsin School of Medicine and Public Health
Objective:
To perform more precise radiologic-pathologic correlation in ex vivo renal cell carcinoma (RCC) for identification of MRI imaging appearance of aggressive pathologic tumor characteristics
Materials and Methods:
A total of 20 patients with large (>7 cm) renal cell carcinoma (RCC) were prospectively recruited for this HIPAA-compliant, IRB-approved study, and written informed consent was obtained. Using preprocedure imaging, a formlabs kidney and tumor mold was 3D printed prior to surgical resection. Once the kidney was surgically removed, it was placed in the mold for orientation and 3 tissue localization clips were inserted into the tumor. The specimen was imaged with noncontrast MRI then taken to surgical pathology. Three 1 cm pathology samples were taken around each tissue localization clip and a pathologic map was generated showing the nuclear grade for that site within the tumor. This was precisely correlated with ex vivo MRI using the clips as landmarks, corresponding 2D ROIs were drawn, and radiomics features were extracted from each site. Nuclear grade 1 - 2 was categorized as low grade, 3 - 4 high. Ten machine learning algorithms were trained using radiomics data from 3D Cor T1 Fast SPGR MR images to identify high versus low grade tumors.
Results:
A total of 18 tumors (14 clear cell) were scanned with MRI (13 men and 5 women, mean age 63 years), 117 ROIs were evaluated with imaging and detailed pathologic evaluation (n=61 low grade, 56 high grade). The support vector machine (SVM) algorithm demonstrated the best performance with an accuracy and AUC of 0.88/0.88, and sensitivity/specificity 83%/92% for identification of high grade tumors on T1 non contrast images. The k-nearest neighbors (knn) algorithm was the next best performer with accuracy and AUC of 0.75, and sensitivity/specificity 83%/67%. The knn algorithm performed best in identifying areas of tumor necrosis, with accuracy and AUC of 0.67, and sensitivity/specificity 58%/75%.
Conclusion:
When precise radiologic-pathologic correlation is performed in heterogeneous RCC, high nuclear grade may have identifiable MRI-based radiomics features.
