Automated Lung Cancer Segmentation Using a Dual-Modality Deep Learning Network with PET and CT Images

Abstract

<h3>Purpose/Objective(s)</h3> To develop an automated lung cancer segmentation method using dual-modality imaging and deep learning, and to perform clinical evaluation of the method. <h3>Materials/Methods</h3> A 3D neural network with dual inputs from diagnostic PET and simulation CT was constructed based on U-Net. The architecture consisted of two parallel convolution paths for independent feature extractions from PET and CT at multiple resolution levels and a single deconvolution path. At each resolution level, the extracted features from the convolution arms were concatenated and fed into the deconvolution path through skip connections. The network was trained/validated/tested by a 3:1:1 split on a dataset of 290 pairs of PET and CT from lung cancer patients treated at our institution, with manual physician contours as the ground truth. The performance of the 3D dual-modality network was compared against that of a CT-only network. evaluated both the manual and the network-produced tumor contours of a randomly selected subset of 20 cases (10 large and 10 small) in a blinded fashion. <h3>Results</h3> The mean Dice similarity coefficient (DSC), Hausdorff Distance (HD), and bi-directional local distance (BLD) comparing the automatic contours versus the ground truth were 0.77 ± 0.12, 7.6 ± 4.7 mm, and 2.9 ± 1.4 mm, and 0.79 ± 0.10, 5.8 ± 3.2 mm, and 2.8 ± 1.5 mm for dual modality inputs, respectively. The stratification method delivered the best results when the model for the large GTV subset (> 25 ml) was trained with GTVs of all sizes (DSC, HD, BLD of 0.85 ± 0.05, 9.5 ± 3.8 mm, and 3.8 ± 1.8 mm), and that for the small GTV subset (< 25 ml) was trained with small GTVs only (DSC, HD, BLD of 0.82 ± 0.08, 3.7 ± 1.7 mm, and 2.2 ± 1.1 mm). From the stratified results, the best combined overall DSC, HD, and BLD were 0.83 ± 0.07, 5.9 ± 2.5 mm, and 2.8 ± 1.4 mm, respectively. In the multi-observer review, on average 91.25% of manual vs. 88.75% of automatic contours were Accepted or Accepted with Modifications. (50% Accepted, 41.25% Accepted w/ Mods and 6.25% Rejected for manual vs. 18.75%, 70%, and 8.75% for automatic), the modifications on the automatic contours were relatively minor with a mean DSC of 0.92 ± 0.04 between the original and modified, comparable to the mean DSC of 0.90 ± 0.04 between all the modified contours and their manual ground truths. <h3>Conclusion</h3> By utilizing an expansive clinical PET and CT image database and a dual-modality architecture, the proposed 3D network with a novel GTV volume-based stratification strategy was able to generate clinically useful lung cancer contours that were quantitatively similar to the ground truth and highly acceptable in physician review.