Siddharth Guha, Abdalla Ibrahim, Qian Wu, Pengfei Geng, Yen Chou, Hao Yang, Jingchen Ma, Lin Lu, Delin Wang, Lawrence H. Schwartz, Chuan-miao Xie, Binsheng Zhao
Abstract
Contrast-enhanced computed tomography scans (CECT) are routinely used in the evaluation of different clinical scenarios, including the detection and characterization of hepatocellular carcinoma (HCC). Quantitative medical image analysis has been an exponentially growing scientific field. A number of studies reported on the effects of variations in the contrast enhancement phase on the reproducibility of quantitative imaging features extracted from CT scans. The identification and labeling of phase enhancement is a time-consuming task, with a current need for an accurate automated labeling algorithm to identify the enhancement phase of CT scans. In this study, we investigated the ability of machine learning algorithms to label the phases in a dataset of 59 HCC patients scanned with a dynamic contrast-enhanced CT protocol.
Introduction
Computed tomography (CT) is routinely used in evaluating liver pathology and provides high diagnostic value. In particular, multi-phase contrast-enhanced or dynamic CT images have been shown to have high sensitivity and specificity in diagnosing primary liver lesions, such as hepatocellular carcinoma (HCC) [1–5]. In fact, dynamic CT plays an essential role in the diagnosis, treatment, and response assessment of hepatocellular carcinoma based on the widely utilized Liver Imaging Reporting and Data System (LI-RADS) [6]. However, not only are there significant differences in HCC lesion detection amongst the various phases of dynamic CT [7, 8], but different phases also provide different diagnostic accuracy for different types of lesions [9, 10].
Materials and method
Data were obtained from two cohorts A, B totaling 135 patients with 2391 scan timepoints analyzed retrospectively (see Figs 1 and 2 for details): [A] dynamic CT scans from 59 patients with 2314 scan timepoints; [B] multiphase CT scans from 76 patients with 77 scans total (one scan of a single phase selected for each patient with one patient with scans of two different phases selected). Scans from Cohort A were from patients diagnosed with HCC (pre-treatment) who were imaged at Sun Yat-sen University Cancer Center using a dynamic CT protocol that took approximately 40 images every 2 seconds following contrast injection.
Results
Each supervised learning model was trained and tested using an input of either mean density from Input A (aorta and portal vein), Input B (aorta and liver parenchyma), or Input C (aorta, portal vein, and liver parenchyma) with the 5-fold cross-validation approach outlined previously. An example of the normalized confusion matrices from a single instance of the cross-validation for the GBDT model using Input C.
Discussion
In order to optimize the automatic classification of dynamic CT scan images into the correct contrast-enhancement phases, we trained, tested, and externally validated five machine learning models using density information from three different ROIs. There were no significant differences in the performance of the various models, and the algorithms achieved relatively high accuracies around the range of 80–90%. The MCC scores of the models when tested in both the original and external datasets also suggest a strong positive correlation between the selected organ density measurements and the contrast-enhancement phase (since the MCC is a contingency matrix method of calculating the Pearson product-moment correlation coefficient [36], the interpretation of the statistic is similar). Thus, there is significant confidence that these models can be generalizable to the broader application of identifying phases during routine dynamic CT imaging.
Conclusions
Simple supervised learning models using only the density from two common anatomical landmarks as an input can accomplish the task of phase identification with high accuracy comparable to more complex models. In the future, these models can be used to develop a fully automated and readily employable phase classifier that can improve the quality of clinical decision-making based on imaging features.
Citation: Guha S, Ibrahim A, Wu Q, Geng P, Chou Y, Yang H, et al. (2024) Machine learning-based identification of contrast-enhancement phase of computed tomography scans. PLoS ONE 19(2): e0294581. https://doi.org/10.1371/journal.pone.0294581
Editor: Shuai Ren, Affiliated Hospital of Nanjing University of Chinese Medicine: Jiangsu Province Academy of Traditional Chinese Medicine, CHINA
Received: April 28, 2023; Accepted: November 4, 2023; Published: February 2, 2024
Copyright: © 2024 Guha et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the paper and its Supporting Information files.
Funding: B.Z. and L.S. received financial support from the National Institutes of Health through grant U01 CA225431 (https://reporter.nih.gov/search/V24b0F7_OU-Zdua-r0RKIQ/project details/10417115) The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Source: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0294581#sec012
