Automated detection and segmentation of small cell lung cancer liver metastases on CT.

Abstract

e13555 Background: Small cell lung cancer (SCLC) is a highly malignant cancer that commonly metastasizes throughout the body (mSCLC) and is molecularly classified based on expression of four transcription regulators (NAPY): NeuroD1, ASCL1. POU2F3, and YAP1. Medical imaging is used for treatment monitoring and patient management. Due to the relatively rare occurrence of this disease, large cohorts of data are not available for training. The purpose of this work is to develop a model for automated detection and segmentation of mSCLC liver metastases by leveraging publicly available data and characterize its output against NAPY subtyping. Methods: Two datasets were used: (1) the Liver Tumor Segmentation Challenge (LiTS), a public dataset containing 130 CT scans and (2) 23 patients with mSCLC imaged at our institution with biopsy-confirmed SCLC metastasis to the liver and available NAPY subtyping. N = 12 mSCLC scans underwent ground truth lesion segmentation by an experienced radiologist (> 10yrs) using ITK-Snap. The LiTS dataset was split to 80/12/8 for training/validation/testing. The mSCLC data was reserved for testing. CT scans were resampled to uniform z-spacing of 1mm and Hounsfield range. A liver organ segmentation model based on 2D UNet and liver lesion segmentation model based on DeepLabV3+ were trained using Fast.ai library. Random flipping, rotation, erasing, and variable CT windowing were applied in each batch. Hyperparameters included cross-entropy loss, Adam optimization, batch size 32, input size [512,512], 1e-4 learning rate, 100 epochs, pre-trained weights from ImageNet. The final liver volume was selected as largest 3D-connected component of output and was used as input to lesion segmentation. Segmentation performance was evaluated by Dice Similar Coefficient (DSC). AI-based lesion segmentations were used to extract 106 shape and texture features (PyRadiomics) for association to NAPY subtypes using logistic regression analysis accounting for the clustered nature of lesion data on the patient-level. Results: The liver model achieved median DSC 0.952 (range (0.926-0.964) and 0.958 (range 0.933-0.969) in the LiTs validation and testing sets, respectively. Lesion segmentation results in validation and testing sets achieved median DSC 0.639 (0-0.848) and 0.523 (0-0.878), respectively. This model demonstrated acceptable generalizabilty in the mSCLC patients from our institution with median Dice coefficient 0.696 (0.102-0.882). Overall, the model achieved 62.1% sensitivity and 88.2% PPV with median 1 FP/image (range 0-4) from 53 ground-truth lesions. Sensitivity analysis showed lesions > 2mm3 had excellent detection performance. AI-based Maximum 2D Diameter and GLCM Difference Entropy were significantly associated with NA vs. PY subtypes across 29 lesions with available outcomes. Conclusions: AI-based models have potential for CT-based disease characterization of liver metastases in mSCLC.