Oral & Poster Presentations Archive

UK & Ireland Prostate Brachytherapy Conference 2022

Abstract number 1 

Auto-segmentation in LDR prostate BT: an analysis of the effective impact of image/segmentation protocol-based domain shift analysis on the performance of deep learning-based prostate segmentation from MRI architecture nn-Unet

Alqarni M*, Jones E**, Mullassery V**, Morris S**, Guerrero Urbano T**, King A*
* King’s College London
** Guy’s and St Thomas’ Hospital, London

PURPOSE: To study the impact of domain shift on prostate MRI auto-segmentation that arises from different acquisition/segmentation protocols with a view to integrate MRI in the BT workflow. 

METHODS: The nn-Unet1 deep learning model was used for prostate auto-segmentation from T2 axial weighted MRIs acquired using 3T Siemens MRI scanners from three publicly available datasets: ProstateX, NIC ISBI 13, I2CVB2-4 (18 subjects from each dataset). Four models were trained: three using data from a single dataset, and a fourth using mixed data from all three datasets. The Dice Similarity Coefficient (DSC) was used to compare the ground truth and model-produced segmentations in three ways: Inter-domain, in which a model was trained on a single dataset and tested on another dataset, Intra-domain, i.e. trained and tested on the same dataset, and mixed domain, i.e. trained on data from all datasets and tested on all datasets. In addition to DSC, the results from the mixed data model were visually assessed by three consultant clinical oncologists from Guy’s and St Thomas’ hospitals using RCR/LIKERT scales. 

RESULTS: Intra-domain performances were significantly (p<0.05) higher than inter-domain performances and noticeable improvement was seen when mixing training data from different datasets. The results show the utility of training models using heterogenous data and that data origin should also be considered as well as scanner vendor as a source of possible domain shift. A Cohen Kappa test showed poor agreement between experts when evaluating the segmentation as “no change” needed but better agreement with “minor change” and “major change” scores. 

CONCLUSION: Auto-segmentation models are sensitive to imaging acquisition and annotation protocols, as well as the well-known effect of scanner vendors. The poor agreement between the spatial overlap scores and the visual assessment scores necessitates the involvement of clinical metrics to further evaluate and tune deep learning models. 

References: 

1. Isensee F, Jaeger PF, Kohl SA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 2021;18(2):203-211. 

2. Armato SG, Huisman H, Drukker K, Hadjiiski L, Kirby JS, Petrick N, et al. PROSTATEx Challenges for computerized classification of prostate lesions from multiparametric magnetic resonance images. Journal of Medical Imaging 2018;5(4):044501. 

3. Clark K, Vendt B, Smith K, Freymann J, Kirby J, Koppel P, et al. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. J Digital Imaging 2013;26(6):1045-1057. 

4. Lemaître G, Martí R, Freixenet J, Vilanova JC, Walker PM, Meriaudeau F. Computer-aided detection and diagnosis for prostate cancer based on mono and multi-parametric MRI: a review. Comput Biol Med 2015;60:8-31. 

Finance-source: The first author is funded through a grant from the Royal Embassy of Saudi Arabia Cultural Bureau.