E1571. External Validation of a Machine Learning Based Algorithm to Differentiate Hepatic Cystadenoma from Benign Liver Cysts
  1. Felipe Furtado; Athinoula A. Martinos Center for Biomedical Imaging; Massachusetts General Hospital
  2. Alvaro Badenes Romero; Athinoula A. Martinos Center for Biomedical Imaging; Massachusetts General Hospital
  3. Leila Mostafavi; Massachusetts General Hospital
  4. Marcelo Queiroz; Hospital das Clinicas HCFMUSP
  5. Mark Anderson; Massachusetts General Hospital
  6. Onofrio Catalano; Athinoula A. Martinos Center for Biomedical Imaging; Massachusetts General Hospital
Historically, imaging has had difficulties differentiating between benign hepatic cysts (BHC) and mucinous cystic neoplasms (MCN), formerly known as biliary cystadenomas or cystadenocarcinomas. Such differentiation is critical because, unlike BHC, MCN has a 20 - 30% chance of malignant transformation. Recently, Hardie et al. proposed a machine learning-derived algorithm to noninvasively classify a liver cyst into BHC or MCN based on CT or MRI imaging features. The objective of this study was to perform an external validation of this diagnostic algorithm in a cohort from a large academic health system with multiple hospitals.

Materials and Methods:
This retrospective observational study evaluated patients with focal liver lesions pathologically confirmed as MCN or BHC between January 2005 and March 2022. Two readers reviewed the contrast-enhanced CT or contrast-enhanced MRI images in consensus and followed the classification system from Hardie et al. to differentiate between MCN and BHC. The algorithm proposes four distinct lesion classes according to the following features: (1) the presence of septations and their relationship to macro-lobulations of the cyst wall, (2) the presence of solid enhancing mural nodules, (3) the number of cystic hepatic lesions. The classification assigned by the readers was then compared to the pathology results, which were considered the ground truth.

The final study cohort was comprised of 172 patients with a median age of 62 years (IQR [51.8, 70.0]), 65.1% women (112), and 34.9% (60) men. Of all patients, 89.0% (153) had benign hepatic cysts (BHC), and the remaining 11.0% (19) had mucinous cystic neoplasms (MCN) on pathology. Imaging was performed via MRI in 51.2% of the cases (88), while the remaining 48.8% (84) were scanned with CT. The median time from imaging to pathology was 28 days (IQR [11.0, 65.3]). Of the 153 benign lesions, 84.3% were class I (129), 14.4% were class II (22), 0.7% were class III (1), and 0.7% were class IV (1). On the other hand, for the 19 MCNs, 15.8% were class I (3), 10.5% were class II (2), 47.4% were class III (9), and 26.3% were class IV (5). When collapsing classes I and II as suggestive of BHC and III and IV as indicative of MCN, the algorithm had an accuracy of 95.9% (95% CI [91.8% - 98.3%]), a positive predictive value of 87.5% (95% CI [61.7% - 98.4%]), a negative predictive value of 96.8% (95% CI [92.7% - 99.0%]), and an area under the receiver operator characteristic curve (AUC) of 0.862 (95% CI [0.760 - 0.964]).

The simple and explainable decision tree discussed in this study is a feasible and accurate tool to aid radiologist decision-making in the cross-sectional evaluation of cystic liver lesions with CT and MRI. This could be further investigated via prospective studies, ideally addressing hard outcomes such as unnecessary biopsy or surgery rates.