Deep learning for automatic tumour segmentation in PET/CT images of patients with head and neck cancers

TL;DR

This paper presents a deep learning-based algorithm using U-Net architecture for automatic tumour segmentation in PET/CT images of head and neck cancer patients, achieving near-oncologist level accuracy.

Contribution

It introduces a CNN-based segmentation method validated on a sizable patient dataset, demonstrating high accuracy and generalizability for clinical use.

Findings

Dice coefficient for PET/CT: 0.75 ± 0.12

Close-to-oncologist level delineation performance

Validated on 55 patient images in total

Abstract

An automatic segmentation algorithm for delineation of the gross tumour volume and pathologic lymph nodes of head and neck cancers in PET/CT images is described. The proposed algorithm is based on a convolutional neural network using the U-Net architecture. Several model hyperparameters were explored and the model performance in terms of the Dice similarity coefficient was validated on images from 15 patients. A separate test set consisting of images from 40 patients was used to assess the generalisability of the algorithm. The performance on the test set showed close-to-oncologist level delineations as measured by the Dice coefficient (CT: $0.65 \pm 0.17$, PET: $0.71 \pm 0.12$, PET/CT: $0.75 \pm 0.12$).