Predicting Treatment Response in Prostate Cancer Patients Based on Multimodal PET/CT For Clinical Decision Support

Sobhan Moazemi, Markus Essler, Thomas Schultz, and Ralph A. Bundschuh
In proceedings of MICCAI Workshop on Multimodal Learning and Fusion Across Scales for Clinical Decision Support, pages 22-35, 2021
 

Abstract

Clinical decision support systems (CDSSs) have gained critical importance in clinical practice and research. Machine learning (ML) and deep learning methods are widely applied in CDSSs to provide diagnostic and prognostic assistance in oncological studies. Taking prostate cancer (PCa) as an example, true segmentation of pathological uptake and prediction of treatment outcome taking advantage of radiomics features extracted from prostate-specific membrane antigen-positron emission tomography/computed tomography (PSMA-PET/CT) were the main objectives of this study. Thus, we aimed at providing an automated clinical decision support tool to assist physicians. To this end, a multi-channel deep neural network inspired by U-Net architecture is trained and fit to automatically segment pathological uptake in multimodal whole-body baseline 68Ga-PSMA-PET/CT scans. Moreover, state-of-the-art ML methods are applied to radiomics features extracted from the predicted U-Net masks to identify responders to 177Lu-PSMA treatment. To investigate the performance of the methods, 2067 pathological hotspots annotated in a retrospective cohort of 100 PCa patients are applied after subdividing to train and test cohorts. For the automated segmentation task, we achieved 0.88 test precision, 0.77 recall, and 0.82 Dice. For predicting responders, we achieved 0.73 area under the curve (AUC), 0.81 sensitivity, and 0.58 specificity on the test cohort. As a result, the facilitated automated decision support tool has shown its potential to serve as an assistant for patient screening for 177Lu-PSMA therapy.

Images

Download Paper

Download Paper

Bibtex

@INPROCEEDINGS{Moazemi:MLCDS2021,
     author = {Moazemi, Sobhan and Essler, Markus and Schultz, Thomas and Bundschuh, Ralph A.},
      pages = {22--35},
      title = {Predicting Treatment Response in Prostate Cancer Patients Based on Multimodal PET/CT For Clinical
               Decision Support},
  booktitle = {MICCAI Workshop on Multimodal Learning and Fusion Across Scales for Clinical Decision Support},
     series = {LNCS},
     volume = {13050},
       year = {2021},
   abstract = {Clinical decision support systems (CDSSs) have gained critical importance in clinical practice and
               research. Machine learning (ML) and deep learning methods are widely applied in CDSSs to provide
               diagnostic and prognostic assistance in oncological studies. Taking prostate cancer (PCa) as an
               example, true segmentation of pathological uptake and prediction of treatment outcome taking
               advantage of radiomics features extracted from prostate-specific membrane antigen-positron emission
               tomography/computed tomography (PSMA-PET/CT) were the main objectives of this study. Thus, we aimed
               at providing an automated clinical decision support tool to assist physicians. To this end, a
               multi-channel deep neural network inspired by U-Net architecture is trained and fit to automatically
               segment pathological uptake in multimodal whole-body baseline 68Ga-PSMA-PET/CT scans. Moreover,
               state-of-the-art ML methods are applied to radiomics features extracted from the predicted U-Net
               masks to identify responders to 177Lu-PSMA treatment. To investigate the performance of the methods,
               2067 pathological hotspots annotated in a retrospective cohort of 100 PCa patients are applied after
               subdividing to train and test cohorts. For the automated segmentation task, we achieved 0.88 test
               precision, 0.77 recall, and 0.82 Dice. For predicting responders, we achieved 0.73 area under the
               curve (AUC), 0.81 sensitivity, and 0.58 specificity on the test cohort. As a result, the facilitated
               automated decision support tool has shown its potential to serve as an assistant for patient
               screening for 177Lu-PSMA therapy.}
}