Discussion of a Simple Method to Generate Descriptive Images Using Predictive ResNet Model Weights and Feature Maps
Destie Provenzano1, Jeffrey Wang2, Sharad Goyal2
1School of Engineering and Applied Science, George Washington University, Washington, DC 20052, USA.
Tomography (Ann Arbor, Mich.)
|March 26, 2025
Summary
Researchers generated simulated MRI images from a predictive model to understand how it identifies cervix cancer recurrence after radiotherapy. This method aids in explaining model decisions and identifying key imaging features for disease course prediction.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Predictive models using Magnetic Resonance Imaging (MRI) can accurately identify cervix tumors likely to recur after radiotherapy (RT).
- Explainability of these models remains a challenge, limiting insight into their decision-making processes.
- This study explored generating simulated images from model features for enhanced model explainability.
Purpose of the Study:
- To investigate the use of model-derived features for generating simulated MRI images.
- To assess the explainability of a Residual Neural Network (ResNet) model for cervix cancer recurrence prediction.
- To determine if simulated images can aid in identifying features predictive of disease course.
Main Methods:
- T2-weighted MRI data from 27 cervix cancer patients treated with RT were analyzed.
- A ResNet model was trained to predict cancer recurrence.
- Feature maps were extracted, combined to create simulated images, and reviewed by a radiation oncologist.
Main Results:
- The ResNet model achieved 93% accuracy in predicting recurrence.
- Generated simulated images successfully mimicked recurrent and non-recurrent cervix tumors.
- A radiation oncologist confirmed simulated images showed aggressive cancer characteristics, including some non-clinically relevant MRI features.
Conclusions:
- A straightforward method generated simulated MRI data resembling cervix cancer recurrence patterns.
- These simulated images can enhance the explainability of predictive models.
- The approach may assist radiologists in identifying prognostic imaging features for disease progression.
