Complete Pathologic Response Prediction by Radiomics Wavelets Features of Unenhanced CT Simulation Images in Locally
Myroslav Lutsyk1, Konstantin Gourevich2, Zohar Keidar2
1Department of Oncology, Rambam Health Care Campus, affiliated with Rappaport Faculty of Medicine, Technion-lsrael Institute of Technology, Haifa, Israel.
Background:
For locally advanced rectal cancer patients a watch-and-wait strategy is an acceptable treatment option in cases of complete tumor response. Clinicians need robust methods of patient selection after neoadjuvant chemoradiation.
Objectives:
To predict pathologic complete response (pCR) using computer vision. To analyze radiomic wavelet transform to predict pCR.
Methods:
Neoadjuvant chemoradiation for patients with locally advanced rectal adenocarcinoma who passed computed tomography (CT)-based simulation procedures were examined. Gross tumor volume was examind on the set of CT simulation images. The volume has been analyzed using radiomics software package with wavelets feature extraction module. Statistical analysis using descriptive statistics and logistic regression was performed was used. For prediction evaluation a multilayer perceptron algorithm and Random Forest model were used.
Results:
In the study 140 patients with II-III stage cancer were included. After a long course of chemoradiation and further surgery the pathology examination showed pCR in 38 (27.1%) of the patients. CT-simulation images of tumor volume were extracted with 850 parameters (119,000 total features). Logistic regression showed high value of wavelet contribution to model. A multilayer perceptron model showed high predictive importance of wavelet. We applied random forest analysis for classifying the texture and predominant features of wavelet parameters. Importance was assigned to wavelets.
Conclusions:
We evaluated the feasibility of using non-diagnostic CT images as a data source for texture analysis combined with wavelets feature analysis for predicting pCR in locally advanced rectal cancer patients. The model performance showed the importance of including wavelets features in radiomics analysis.
