2475. MRI-Based Three-Dimensional Multichannel Convolutional Neural Network for the Differentiation of Benign from Malignant Solid Renal Masses
Authors * Denotes Presenting Author
  1. Peter Wawrzyn *; University of Wisconsin-Madison
  2. Ruben Ngnitewe Massa; University of Wisconsin-Madison
  3. Jamal Gardezi; University of Wisconsin-Madison
  4. Andrew Wentland; University of Wisconsin-Madison
MRI is often used in the evaluation of solid renal masses. Such solid renal masses include malignant entities, such as renal cell carcinoma (RCC) subtypes, as well as benign lesions, such as oncocytomas and angiomyolipomas (AMLs). Renal masses are often biopsied for further characterization. A non-invasive diagnostic tool to aid in tumor classification would be useful clinically. Artificial intelligence-based techniques for evaluating renal masses have been explored in recent years. However, many studies employing these techniques have limited data sets lacking the range of pathology typically seen clinically or have developed models that fail to predict pathology in such a way that would mitigate the need for biopsy. The purpose of this study was to develop a 3D multichannel convolutional neural network (CNN) to determine whether a tumor is benign or malignant. Previous work in our lab has led to the curation of a balanced and representative dataset of MRI scans of solid renal masses.

Materials and Methods:
An MRI dataset of 175 solid renal masses was curated. This cohort included 111 malignant tumors (73 clear cell RCC, 25 papillary RCC, 13 chromophobe RCC) and 64 benign tumors (48 AMLs, 16 oncocytomas). Masses were segmented with a bounding box independently for four tissue contrast weightings, including T2-weighted images as well as postcontrast T1-weighted images in the arterial, 30-second delayed, and 4-minute delayed phases. The data were divided 80/20 into training and testing sets. Five-fold cross-validation was employed on the training set. The model architecture consisted of four convolutional layers with max pooling and ReLU activation followed by three dense fully connected layers, as previously described. Training of the model occurred over 30 epochs to minimize the binary cross-entropy loss of the model. The model parameters that achieved the highest validation accuracy for each fold were then selected for testing. Summative metrics included accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, Matthew’s correlation coefficient (MCC), Youden’s J statistic (J), and area under the curve (AUC).

The best performing deep learning model was used all four tissue contrast weightings as input channels, yielding accuracy = 0.80, sensitivity = 1.00, specificity = 0.46, PPV = 0.76, NPV = 1.00, F1 score = 0.86, MCC = 0.59, J = 0.46, and AUC = 0.59. Accuracies were lower when the model was trained with each tissue contrast weighting independently rather than via a multichannel approach, yielding accuracies of 0.76, 0.75, 0.69, and 0.75 for T2, T1 arterial, and T1 30-second and T1 4-minute delayed images, respectively.

A CNN model trained from MRI data can differentiate benign from malignant solid renal masses with a high degree of accuracy. Performance is improved with a multichannel approach that utilizes multiple MR tissue contrast weightings.