ESTRO 2024 - Abstract Book

S3096

Physics - Autosegmentation

ESTRO 2024

1 Umeå University, Department of Radiation Sciences, Umeå, Sweden. 2 University of Szeged, Department of Radiology, Szeged, Hungary. 3 Skåne University Hospital, Department of Haematology, Oncology and Radiation Physics, Lund, Sweden

Purpose/Objective:

In the context of prostate cancer treatment, radiotherapy has traditionally been delivered with a homogeneous dose to the entire prostate. Recent studies have demonstrated the efficacy of a local dose escalation to a subvolume of the prostate, so called focal boost. This highlights the importance of delineating the intraprostatic organ-at-risk (OAR) urethra, to minimize side-effects [1]. In the diagnostic phase, the global PI-RADS standard is used to assess prostatic lesions on MRI with different dominant sequences depending on their zonal location [2]. These anatomical zones of the prostate have different histological features and are defined as the peripheral zone (PZ), central zone (CZ), transitional zone (TZ) and anterior fibromuscular stroma (AFS). The unique characteristics of each zone could potentially inform individualized treatment and risk stratification of patients, making treatment zonal dependent instead of zonal agnostic. Manual segmentation of the prostate, urethra, and its anatomical zones on MRI are tedious and time-consuming and standardized methods are not appropriate due to individual variations between patients. Therefore, the development of an individualized, automatic method to segment the prostate, urethra as well as its internal zones is potentially highly relevant in present medical practice, with implications both for treatment and diagnosis.

In this work, we evaluate a deep learning algorithm on a segmentation dataset including our manual delineations of the prostate, urethra, and prostatic zones.

Material/Methods:

Manual segmentations of the urethra and prostatic zones were performed on a cohort of 200 randomly selected prostate mp-MRIs from the public PROSTATEx dataset [3]. The segmentations were performed by two experienced radiologists (>1000 and >500 interpreted prostate MRIs) in collaboration with three junior colleagues (<100 interpreted prostate MRIs each). Segmentations by the junior colleagues were checked and, if necessary, corrected by the experienced radiologists. The urethra was delineated as a circle with a 6 mm diameter in each intra-prostatic slice, resembling a radiotherapy OAR-delineation [1]. All segmentations were performed using Slicer 3D (version 5.2.2). A multi-professional team performed minor adjustment, removing isolated pixels and harmonizing segmentations between slices. Subsequently, all segmentations were submitted for final approval from the two most experienced radiologists.

The dataset was divided into a training set (n=160) and a test set (n=40). The test set were segmented by both experienced radiologists, resulting in 240 segmentations for 200 patients.

We trained a 3D nnUNet model [4] with five-fold cross-validation using the T2w images and segmentations from the training dataset.

The probability maps of the model outputs were used in a postprocessing step to delineate the urethra within the prostate boundary in the desired circular structure. For all slices the centre point of the urethra was identified based on the average indices of pixels with a probability greater than a given threshold value and within 90% of the slice wise maximum probability. Any slices, between two other slices with a delineated urethra, without indices fulfilling the above criteria were linearly interpolated.