Segmentation of malignant lesions on [18F]FDG PET/CT images using deep learning techniques

Abstract

Cancer is one of the leading causes of death worldwide, and that can be prevented through early detection and effective treatment of malignancies. Hence, tumour identification and segmentation take a vital part in the early detection of lesions, as well as in radiotherapy procedures, and surgical planning. At present, manual segmentation is the gold standard, which remains a daunting challenge as it is time-consuming, labour intensive, tedious, and highly subjective, since introduces inter- and intra-operator variability. Thereby, the paramount purpose of this internship was to apply, optimise and assess the feasibility of deep learning-based techniques for the automatic identification and segmentation of malignant lesions in whole-body [18F]FDG PET/CT images. Hence, three different datasets were used to train different networks: CT images for spleen segmentation; [18F]FDG PET/CT with suspected malignant lesions; and lesions suggestive of lymphoma on whole-body PET images. Subsequently, a 3D U-net architecture was developed and optimised for automatic identification and segmentation of the objects of interest. Due to GPU computational capacity limitations, several approaches needed to be implemented for the 3D U-net training process and inference testing. The Dice coefficient (DC) was used as an overlap measure between the ground truths and the resulting segmentations. The first dataset achieved the highest median DC of 0.57 using the network with the transfer learning method and a CT intensity normalisation of [-250; 250]. Regarding the second dataset, the median DC obtained was 0.28, when the 3D U-net was trained with patches of 48×48×48. Finally, the third dataset achieved a median DC of 0.41, whereas the patch size was 64×64×64 voxels on a U-net configuration with one less layer. In conclusion, fully automatic segmentation methods based on deep learning techniques for lesion identification and segmentation need clinical supervision for verification and adjustments. For now, this method is unacceptable to use in clinical practice alone, since it is not robust enough.

Date of Award	23 Feb 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Francisco Paulo Marques de Oliveira (Supervisor) & Cláudia Santos Constantino (Co-Supervisor)