3554 - A Simple Method to Generate Descriptive Images for Recurrent Cervix Cancer Using Predictive ResNet Model Weights and Feature Maps

D. Provenzano¹, M. Loew², S. Goyal³, and Y. J. Rao³; ¹Biomedical Engineering, George Washington University School of Engineering and Applied Science, Washington, DC, ²Medical Oncology, George Washington University School of Medicine and Health Sciences, Washington, DC, ³Radiation Oncology, George Washington University School of Medicine and Health Sciences, Washington, DC

Purpose/Objective(s): We previously demonstrated that a Residual Neural Network (ResNet) can use magnetic resonance imaging (MRI) data to identify cervix tumors at increased risk of recurrence after radiotherapy (RT). In addition to being used as classifiers, ResNets can be used as feature extractors. In this study we explored whether features extracted from a highly accurate predictive model for recurrent cervix cancer after RT could be used also to generate simulated MRI data indicative of recurrent cervix cancer. Materials/

Methods: Twenty-seven women with cervix cancer who received radiotherapy treatment were identified in the The Cancer Genome Atlas Cervical Squamous Cell Carcinoma and Endocervical Adenocarcinoma Collection (TCGA-CESC). Corresponding T2W MRI data for them were collected from the Cancer Imaging Archive (TCIA). Simulated images were generated according to our algorithm as follows: [A] A ResNet model was trained to identify recurrent cervix cancer with high accuracy. [B] The model was evaluated on T2W MRI data from subjects with and without recurrent tumors to obtain corresponding feature maps. [C] The best predictive feature maps from each image were determined based on ResNet predictions. [D] The best predictive features from each feature map were then stacked to create a new simulated image. [E] The final generated simulated image was passed back through the initial algorithm to identify likelihood of recurrence.
Results: Predictive feature maps from the ResNet model (93% accuracy) were used to generate simulated images. Simulated images passed through the model were identified as recurrent and non-recurrent cervix tumors after radiotherapy respectively. A radiation oncologist identified the simulated images as cervix tumors with characteristics of aggressive cervical cancer.
Conclusion: This simple method was able to generate simulated MRI data that mimicked recurrent and non-recurrent cervix cancer tumor images. Future refinement of this algorithm will focus on generating multiple types of highly predictive simulated images to identify potential explainable features for algorithm performance. These generated images could be useful for evaluating explainability of predictive models and to assist radiologists with identification of features likely to predict disease course.