Abstracts

RETURN TO ABSTRACT LISTING


2371. A Sum-Product Network Model for Classification of Renal Masses Using CECT-Based Radiomic Features
Authors * Denotes Presenting Author
  1. Priya Ganapathy *; Utopia Compression Corporaton
  2. Lakshmi Priya Rangaraju; Utopia Compression Corporaton
  3. Bino Varghese; University of Southern California
  4. Gautam Kunapuli; Verisk Analytics
  5. Darryl Hwang; University of Southern California
  6. Bhushan Desai; University of Southern California
  7. Vinay Duddalwar; University of Southern California
Objective:
To evaluate a novel artificial intelligence (AI) engine called sum-product network (SPN) models for use in the imaging evaluation of renal masses using contrast-enhanced computed tomography (CECT) based radiomic features. In comparison to traditional deep learning models, SPN-based deep learner models are interpretable and computationally efficient with comparable classification accuracy.

Materials and Methods:
In this IRB-approved, HIPAA-compliant, retrospective study, 3-dimensional tumor volumes were manually segmented from CECT images of 140 patients with predominantly solid, non-macroscopic fat-containing renal tumors (a total of 97 malignant and 43 benign cases). From the segmented region of interest of each lesion, 10 Haralick texture features from 4 phases of CECT images were extracted. We then implemented a Mixed SPN (MSPN) model with these extracted continuous and discrete radiomic features as inputs using the SPFlow python libraries. MSPN implements a K-means clustering approach to partition the training dataset and further uses Renyi Maximum Correlation Coefficient (RDC) to transform and compare dependency between hybrid features representing a given partitioned dataset. Resultant SPN models are tree-like models, with each radiomic feature and the class labels in the pool representing a leaf/terminal node. The post-resected tumor pathology served as a gold standard to evaluate our method.

Results:
We compared MSPN model to conventional (Naïve Bayes, support vector machines, decision-trees, ensemble methods such as bagging, boosting and random forests) and state-of-the-art (relational functional gradient boosting (RFGB) and artificial neural nets) machine-learning methods. Our MSPN model (ROC-AUC = 0.86) exhibited statistically significant performance (p < 0.01, at 99% confidence interval) compared to other approaches (0.63-0.83).

Conclusion:
Our result demonstrates that SPN is a promising technique for classification of renal masses using CECT-based radiomic features. By adding relevant clinical features or even raw radiology images as inputs, SPN may produce greater performance gains on both binary and multi-class renal mass classification problems. SPN results can be viewed graphically along with automatically-generated textual explanations which are needed for translation into clinical decision-support systems and increased end-user acceptance.