ARRS 2022 Abstracts

RETURN TO ABSTRACT LISTING


E1169. Radiomics With Novel Hybrid Machine/Deep Learning: Towards Improving Bladder Cancer Staging and Treatment
Authors
  1. Min Kong; Mayo Clinic
  2. Suryadipto Sarkar; Arizona State University
  3. Haidar Abdul-Muhsin; Mayo Clinic
  4. Aidan McGirr; Arizona State University
  5. Teresa Wu; Arizona State University
  6. Parminder Singh; Mayo Clinic
  7. Alvin Silva; Mayo Clinic
Objective:
Accurate early bladder cancer staging is important as it determines the mode of initial treatment. Non-muscle invasive bladder cancer (NMIBC) can be treated with transurethral resection, whereas muscle invasive bladder cancer (MIBC) requires neoadjuvant chemotherapy with subsequent cystectomy as indicated. Interpretation of contrast-enhanced CT (CECT) by radiologists does not accurately predict pathologic staging of bladder cancer, with estimated accuracy of 49%. Our project aims to improve the accuracy of bladder cancer staging by CECT using a novel hybrid machine/deep learning model.

Materials and Methods:
We retrospectively identified the patients who were diagnosed with/treated for bladder cancer at our tertiary care institution between 1998 and 2020. Pre-operative CECT images and corresponding pathologic staging data were available for each patient. CECTs with no visible lesion or with significant metal artifacts were excluded. Using ImageJ software, each pathology-proven malignant lesion was manually segmented. The lesions are categorized into NMIBC and MIBC groups. The segmented lesions were analyzed utilizing our novel hybrid machine/deep learning model wherein the deep learning module was employed for feature extraction from the raw images, and the extracted features were analyzed using machine learning classifiers. Employing a hybrid deep/machine learning model ensures to capture detailed representation of the images and addresses the problem of overfitting on a small dataset. Seven machine learning classifier models were implemented including: decision tree, neural network, random forest, support vector machine, discriminant analysis, naïve Bayes, and k-nearest neighborhood. Six metrics are collected to show the model’s performance: accuracy, sensitivity, specificity, precision, F1-score, and AUC.

Results:
Twenty-four NMIBC and 41 MIBC lesions met our inclusion criteria. Our hybrid machine/deep learning model showed outstanding performance in predicting NMIBC and MIBC lesions. All seven machine learning classifiers demonstrated superior accuracy in predicting bladder cancer stage (82.4–94.7%) compared to unaided interpretation by radiologists (49%) per Tritschler et el. Among the seven classifiers, the decision tree classifier demonstrated the best performance: accuracy 94.7%, precision 88.7%, AUC 96.9%, F1-score 90.5%, sensitivity 91.6%, and specificity 90%.

Conclusion:
Our hybrid machine/deep learning model showed superior accuracy in predicting bladder cancer staging compared to unaided interpretation by radiologists. This radiomics-assisted interpretation of CECT by the radiologists will facilitate appropriate clinical management of the patients with bladder cancer, ultimately improving patient outcomes.