A Deep Learning model to segment liver metastases on CT images acquired at different time-points during chemotherapy

Abstract

The aim of this study is to present a fully automatic deep learning algorithm to segment liver Colorectal cancer metastases (lmCRC) on CT images, based on a U-Net structure, comparing nets with and without the transfer learning approach. This is a bi-centric study, enrolling patients who underwent CT exam before (baseline) and after first-line therapy (TP1). Patients were divided into training (using a portion of baseline sequences from both centers) to train the DL model, and two validation sets: one with baseline (valB), and one with TP1 (valTP1) sequences. The reference standard for the automatic segmentations was defined by the manual segmentations performed by an experienced radiologist on the portal phase of the baseline and TP1 CT exam. The best performing model obtained Dice Similarity Coefficient (DSC) of $0.68\pm 0.24$, Precision (Pr) of $0.74\pm 0.27$, Recall (Re) of $0.73\pm 0.26$, Detection Rate (DR) of 93% on the valB, and DSC of $0.61\pm 0.28$, Pr of $0.68\pm 0.31$, Re of $0.65\pm 0.29$ and DR of 88% on the valTP1. These encouraging results, if confirmed on larger dataset, might provide a reliable and robust tool that can be used as first step of future radiomics analyses aimed at predicting response to therapy, improving the management of lmCRC patients.