ESTRO 2020 Abstract book

S1028 ESTRO 2020

Conclusion Four commercially available MAR algorithms were evaluated and compared to a novel stereoscopic technique, AMPP, using an anthropomorphic phantom. The major CT vendors’ solutions evaluated can provide qualitative improvement in artifacts caused by the dental implants, but not as well as the in-house technique developed. AMPP was also evaluated for robustness and proved to perform similarly in spite of different imaging parameters and scanners used. PO-1755 Use of ‘Jigsaw puzzles’ to train convolutional neural networks for segmentation with limited data E. Henderson 1 , E. Vasquez Osorio 1 , M. Van Herk 1 , A. Green 1 1 University of Manchester, Division of Cancer Science, Manchester, United Kingdom Purpose or Objective Segmentation is a key step in the radiotherapy pathway, and is usually done manually by clinicians. Recently, convolutional neural networks (CNNs) have been proposed as a method to automate segmentation of organs at risk (OARs) for radiotherapy. However, current methods for training CNNs require large amounts of training data (hundreds of cases). In this work, a CNN for 3D segmentation of Head and Neck (HN) CT scans is trained using transfer learning from an unsupervised task; in this case the identification of random 3D ‘jigsaw pieces’ extracted out of HN CT scans. Material and Methods A small dataset of 68 HN CT scans from the open TCIA data (https://www.cancerimagingarchive.net) was used to develop a 3D U-Net to segment OARs in the HN. A network resembling the down-arm of the 3D U-Net was first trained to learn relevant features in HN CTs by solving a simpler problem: to identify the jigsaw created by randomly shuffling permutations of small image pieces (here 32 3 pixels) (Fig. 1). We trained the network with 10 selected permutations (from 9! possibilities) per image on the entire dataset, using a stochastic gradient descent optimiser on ‘categorical cross-entropy loss’. The final network for segmentation was assembled by removing the output layers of the jigsaw network and attaching an un-initialised up-arm to complete a full U-Net for 3D segmentation (Fig. 1). In this process, known as transfer learning, the weights from the jigsaw network are used as a starting point. The dataset was split 8:1:1 for training, validation and testing of the final network. We compare segmentations from models with and without jigsaw weight initialisation (i.e., a U-Net trained from scratch), evaluating a DICE coefficient and distance to agreement for the skin , left and right parotids, brainstem and the cervical section of the spinal cord.

Results Training the up-arm networks takes 20 minutes on a Tesla V100, and segmentation takes only 5 seconds. The regular U-Net without jigsaw weight initialisation fails to converge. The transfer learned network has similar performance as networks trained on much larger datasets (Fig. 2).

Conclusion We demonstrated a novel method for training a 3D segmentation CNN using a two-stage process based on a pretext task (identifying jigsaws) and transfer learning. By using transfer learning, excellent results are possible with a small data set; allowing easier development of models for complex tasks such as 3D segmentation. We believe this style of transfer learning has massive potential for applications in medical imaging. Even better accuracy shall be attained through fine-tuning and a sensible method to initialise the up-arm. In future, jigsaw identification will be performed with larger numbers of permutations as this has been shown to increase transfer learning effectiveness.