ARRS 2022 Abstracts


1230. Using Machine Learning to Develop a Classification System to Differentiate Mucinous Cystic Neoplasm of the Liver and Benign Hepatic Cyst
Authors * Denotes Presenting Author
  1. Andrew Hardie; Medical University of South Carolina
  2. Madison Kocher *; Medical University of South Carolina
  3. Jordan Chamberlin; Medical University of South Carolina
  4. Robert Petrocelli; NYU Langone Medical Center
  5. James Boyum; Mayo Clinic
  6. Kedar Sharbidre; University of Alabama at Birmington
  7. Mark Kovacs; Medical University of South Carolina
Although hepatic cystic lesions are common and often benign, cystic neoplasms such as mucinous cystic neoplasms (MCN) can occur in up to 5% of cases and can be misdiagnosed as benign hepatic cysts (BHC). Prior studies have revealed the value of identifying the presence of any septations arising away from an external macro-lobulation (correlating with MCN) as opposed to arising only at an external macro-lobulation (correlating with BHC). The purpose of this study was to use machine learning and a multicenter study design to develop and assess the performance of a novel classification system based on these prior observations for predicting whether a hepatic cystic lesion represents an MCN or BHC.

Materials and Methods:
A multicenter cohort study identified 154 surgically resected hepatic cystic lesions that were pathologic confirmed as MCN (43) or BHC (111). Readers at each institution recorded imaging features previously identified as potential differentiating features from prior publications including lesion size, the presence or absence of a solid enhancing nodule, calcifications, septations, thick septations (3 mm or greater), all septations arising from an external macro-lobulation of the lesion, and whether there were other cystic lesions present in the liver. The contribution of each of these features to differentiating MCN from BHC was assessed by univariate, multivariate and machine learning (Random-Forrest) analyses to develop an optimal classification system.

Several imaging features demonstrated statistical significance on univariate or multivariate analysis. However, only three imaging features were found by machine learning to significantly contribute to a potential classification system: 1) solid enhancing nodule, 2) all septations arising from an external macro-lobulation, and 3) if the lesion was a solitary cystic liver lesion. The optimal classification test-set was 93.5% accurate (95% CI 90.2-97.8) for differentiating MCN from BHC. From this data, an optimized diagnostic flow chart and four-part classification schema could be produced: class 1 lesions represented BHC in 97% of cases (63 % of total lesions); class 2 BHC 89% (8% of lesions); class 3 MCN 88% (27%); and class 4 MCN 100% (2%).

This multi-center study was able to use machine learning to develop a highly accurate classification system for differentiation of hepatic MCN from BHC. This system is likely useful in risk-stratifying cystic hepatic lesions and has practical implications as it could be readily applied to clinical practice.