Kristina I Ringe1, Van Dai Vo Chieu2, Frank Wacker2, Henrike Lenzen3,4, Michael P Manns4, Christian Hundt5, Bertil Schmidt6, Hinrich B Winther2. 1. Department of Diagnostic and Interventional Radiology, Hannover Medical School, Carl-Neuberg Str. 1, 30625, Hannover, Germany. ringe.kristina@mh-hannover.de. 2. Department of Diagnostic and Interventional Radiology, Hannover Medical School, Carl-Neuberg Str. 1, 30625, Hannover, Germany. 3. Department of Gastroenterology and Hepatology, University Hospital Essen, University of Duisburg-Essen, Hufelandstr. 55, 45122, Essen, Germany. 4. Department of Gastroenterology, Hepatology and Endocrinology, Hannover Medical School, Carl-Neuberg Str. 1, 30625, Hannover, Germany. 5. NVIDIA AI Tech Center, Interdisciplinary Centre for Security, Reliability and Trust, Université du Luxembourg, 29, avenue JF Kennedy, L-1855, Luxembourg, Luxembourg. 6. Institute for Computer Science, Johannes Gutenberg University, Saarstraße 21, 55122, Mainz, Germany.
Abstract
OBJECTIVES: To develop and evaluate a deep learning algorithm for fully automated detection of primary sclerosing cholangitis (PSC)-compatible cholangiographic changes on three-dimensional magnetic resonance cholangiopancreatography (3D-MRCP) images. METHODS: The datasets of 428 patients (n = 205 with confirmed diagnosis of PSC; n = 223 non-PSC patients) referred for MRI including MRCP were included in this retrospective IRB-approved study. Datasets were randomly assigned to a training (n = 386) and a validation group (n = 42). For each case, 20 uniformly distributed axial MRCP rotations and a subsequent maximum intensity projection (MIP) were calculated, resulting in a training database of 7720 images and a validation database of 840 images. Then, a pre-trained Inception ResNet was implemented which was conclusively fine-tuned (learning rate 10-3). RESULTS: Applying an ensemble strategy (by binning of the 20 axial projections), the mean absolute error (MAE) of the developed deep learning algorithm for detection of PSC-compatible cholangiographic changes was lowered from 21 to 7.1%. Sensitivity, specificity, positive predictive (PPV), and negative predictive value (NPV) for detection of these changes were 95.0%, 90.9%, 90.5%, and 95.2% respectively. CONCLUSIONS: The results of this study demonstrate the feasibility of transfer learning in combination with extensive image augmentation to detect PSC-compatible cholangiographic changes on 3D-MRCP images with a high sensitivity and a low MAE. Further validation with more and multicentric data is now desirable, as it is known that neural networks tend to overfit the characteristics of the dataset. KEY POINTS: • The described machine learning algorithm is able to detect PSC-compatible cholangiographic changes on 3D-MRCP images with high accuracy. • The generation of 2D projections from 3D datasets enabled the implementation of an ensemble strategy to boost inference performance.
OBJECTIVES: To develop and evaluate a deep learning algorithm for fully automated detection of primary sclerosing cholangitis (PSC)-compatible cholangiographic changes on three-dimensional magnetic resonance cholangiopancreatography (3D-MRCP) images. METHODS: The datasets of 428 patients (n = 205 with confirmed diagnosis of PSC; n = 223 non-PSC patients) referred for MRI including MRCP were included in this retrospective IRB-approved study. Datasets were randomly assigned to a training (n = 386) and a validation group (n = 42). For each case, 20 uniformly distributed axial MRCP rotations and a subsequent maximum intensity projection (MIP) were calculated, resulting in a training database of 7720 images and a validation database of 840 images. Then, a pre-trained Inception ResNet was implemented which was conclusively fine-tuned (learning rate 10-3). RESULTS: Applying an ensemble strategy (by binning of the 20 axial projections), the mean absolute error (MAE) of the developed deep learning algorithm for detection of PSC-compatible cholangiographic changes was lowered from 21 to 7.1%. Sensitivity, specificity, positive predictive (PPV), and negative predictive value (NPV) for detection of these changes were 95.0%, 90.9%, 90.5%, and 95.2% respectively. CONCLUSIONS: The results of this study demonstrate the feasibility of transfer learning in combination with extensive image augmentation to detect PSC-compatible cholangiographic changes on 3D-MRCP images with a high sensitivity and a low MAE. Further validation with more and multicentric data is now desirable, as it is known that neural networks tend to overfit the characteristics of the dataset. KEY POINTS: • The described machine learning algorithm is able to detect PSC-compatible cholangiographic changes on 3D-MRCP images with high accuracy. • The generation of 2D projections from 3D datasets enabled the implementation of an ensemble strategy to boost inference performance.
Entities:
Keywords:
Cholangiography; Deep learning; Machine learning; Sclerosing cholangitis