**1. Introduction**

Blunt spleen injuries account for up to half of all abdominal solid organ injuries. Common causes include road traffic accidents, falls, physical assaults, and sports-related injuries. Multiphasic contrast-enhanced computed tomography (CT) is the standard noninvasive diagnostic tool for injury evaluation of blunt spleen injuries [1], which include lacerations, subcapsular or parenchymal hematomas, active hemorrhage, and vascular injuries. The type and severity of spleen injuries are commonly described based on the Abbreviated Injury Scale (AIS) or the American Association for Trauma (AAST) Organ Injury Scale (OIS). Currently, detection and classification of spleen injuries rely on the manual review of radiologists. This manual process is not only inefficient but also subject to variability based on the reviewer [1,2].

Many computer-assisted diagnosis (CAD) systems have been developed to detect, locate, and assess potential anomalies or injuries to aid radiologists in the diagnostic process. Detection of pathology in the chest, breast, and colon has been the main focus of previous CAD studies [3]. Other extant CAD systems include those that target the brain, liver, skeletal, and vascular systems [3–5]. Although there have not been previous studies on CAD systems for the spleen, an automated method for localizing and segmenting the spleen [6] was previously developed by the co-authors of this study. This method can be

**Citation:** Wang, J.; Wood, A.; Gao, C.; Najarian, K.; Gryak, J. Automated Spleen Injury Detection Using 3D Active Contours and Machine Learning. *Entropy* **2021**, *23*, 382. https://doi.org/10.3390/e23040382

Academic Editor: Amelia Carolina Sparavigna

Received: 15 February 2021 Accepted: 22 March 2021 Published: 24 March 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

utilized to segmen<sup>t</sup> the region of interest in a CT volume, a prerequisite step to performing spleen injury detection.

Machine learning techniques, including Support Vector Machines (SVM), random forest (RF), logistic regression (LR), and deep learning methods, have been widely applied for analysis of medical images [5]. A critical step in application of machine learning to medical image analysis is the extraction and representation of features salient to the classification or detection task at hand. Different types of features and feature extraction methods have been employed based on the anomaly of interest. Common features used include histogrambased [7,8], shape-based [7,9–11], texture-based [12–14], region-based [10,15], and bag-ofwords features [15], among others.

In this paper, we propose a supervised classification scheme to discriminate lacerated spleens from healthy controls, a schematic diagram that is presented in Figure 1. Lacerations were chosen for study as they are major types of blunt spleen injury that can be readily observed from contrast-enhanced CT, appearing as linear or branching regions extending from the capsular surface of the spleen and often disrupting the smooth splenic contour [1]. CT scans from patients experiencing traumatic injuries were collected from the Michigan Medicine and the Crash Injury Research Engineering Network (CIREN) dataset [16]. Healthy and lacerated spleens within CT scans from 99 patients were automatically segmented using a previously developed method [6]. From the segmented spleen region, various features were extracted: statistical histogram-based features including Rényi entropy; shape-based feature including fractal dimension [17], whose generalized version is directly related to Rényi entropy [18]; and texture-based features. The performance of five machine learning models: RF, naive Bayes, SVM, *k*-nearest neighbors (*k*-NN) ensemble, and subspace discriminant ensemble, were trained using 5-fold cross-validation. On a distinct test set, RF was the best performing classifier, achieving an Area Under the receiver operating characteristic Curve (AUC) of 0.91 and F1 score of 0.80. This study demonstrates the potential for such an automated injury assessment method to reduce physician workload and improve patient outcomes by enabling faster injury triage.

**Figure 1.** A schematic diagram of the proposed method.

#### **2. Materials and Methods**
