Session Item

ILD can predispose patients to high risk of pulmonary complications or even death following high dose radiation therapy.   Unfortunately, not all patients with ILD are known at the time of radiation treatment decision. Automated methods to detect ILD on diagnostic and/or planning CTs would provide multiple checks to ensure patient safety, but requires high quality lung segmentation of both patients with and without ILD as a prerequisite.

Using a training set (TRN) of 214 radiation planning  computed tomographic (CT) images from NSCLC patients including cases with and without ILD, a convolutional neural net (CNN) was trained to automatically segment the lungs within these scans.  We trained a CNN based on the U-Net topology but using 3D convolution filters in place of the traditional 2D. Due to memory limitations all images resampled to 256x256x128. After training, two validation datasets were generated composed of radiation treatment planning CTs from NSCLC patients (VAL1, n=24) and a set of diagnostic thoracic CT images (VAL2, n=100[CM1] ).  All patients in VAL2 were further labeled by the same radiologist as to whether ILD was radiographically present or absent.  Test characteristics of the CNN were calculated using the Dice metric on VAL1, and qualitative inspection on VAL2.

After training, the CNN demonstrated Dice of 0.96 on VAL1 and strong qualitative agreement on VAL2, uniquely demonstrating that a CNN can be trained on RT planning CTs to segment both planning and diagnostic imaging. The CNN takes on average 4.5 seconds to segment a novel image with roughly half that time dedicated to reading the image from disk. ILD was present in 20% of cases in VAL2.  

A CNN has been developed that can rapidly segment radiation treatment planning CT or diagnostic CTs to enable downstream automation of a system to identify patients at high risk of having pre-existing ILD.  After prospective validation, this tool and similar tools could be incorporated into radiation treatment planning systems to automatically alert clinicians about high-risk patients that might have proceeded to treatment planning without having pre-existing ILD identified.