Jake Kendrick | Final Masters Presentation | Medical Physics | The University of Western Australia
Automatic Segmentation of Metastatic Prostate Lesions in PET/CT Images Using Convolutional Neural Networks
Supervisors:
Adj/Prof Martin A. Ebert
Dr. Ghulam M. Hassan
Dr. Pejman Rowshan Farzad
=========================
For more information visit our
website: [ Ссылка ]
Weblog: [ Ссылка ]
Facebook: [ Ссылка ]
Instagram: [ Ссылка ]
Twitter: @MedicalUwa
=========================
Abstract
Introduction: The manual segmentation of metastatic prostate lesions in 68Ga-PSMA PET/CT images is a time consuming and laborious process when any given patient may have tens or hundreds of lesions. Developing an automated framework for this process could reduce its time-consuming nature, minimise the inter-observer variability inherent in manual delineations, as well as facilitate efficient treatment planning for patients undergoing radionuclide therapy. This research aims to utilise deep learning techniques to develop an automated segmentation model in whole-body 68Ga-PSMA PET/CT images.
Methods: A local dataset of 336 patients with metastatic prostate cancer, who have had whole-body 68Ga-PSMA PET/CT assessments, was analysed. PSMA-avid lesions were manually segmented by a clinical expert to serve as the ground truth data. This data was subsequently used to train a Convolutional Neural Network (CNN). The Dice Similarity Coefficient (DSC), measuring the level of overlap between model prediction and the ground truth, was chosen to be the primary metric of model evaluation. Secondary metrics, including the sensitivity, specificity, precision, and accuracy score were also calculated.
Results: All models trained show a tendency to perform well on training data, but do not generalise well to previously unseen test cases. The best performing model was trained using the combined multi-channel patient PET/CT images with standard soft tissue Hounsfield Unit (HU) window and level applied to the CT channel, achieving a mean DSC of 0.350 0.308 (median = 0.333). This result is inferior to recent reports in the literature which, to date, have been limited in their investigation to just metastatic prostate lesions in the pelvic area. Training models with each modality individually, i.e., with CT images or PET images alone, diminishes model performance. The best CT-only model and PET-only model achieved mean DSC’s of 0.028 0.1 and 0.336 0.308, respectively.
Conclusion: This novel research undertook an exploratory work on the automated segmentation of metastatic prostate lesions in 68Ga-PSMA PET/CT images. The results presented demonstrate that further experimentation is required, either in the form of pre-processing steps or CNN architecture optimisation, to generate a model with a clinically viable segmentation accuracy.
