DenseNet-based transfer learning for LV shape Classification: Introducing a novel information fusion and data augmentation using statistical Shape/Color modeling

Abstract

PurposeMyocardial infarction (MI) causes the left ventricle (LV) remodeling. Statistical shape modeling (SSM) can provide valuable and reliable information about the changes in the shape and functional patterns of the LV after MI. On the other hand, classification and detection tasks using transfer learning by popular pre-trained convolutional neural networks (CNNs) using a large dataset recently presented considerable results. The current study introduces an innovative hybrid approach for MI detection that combines the benefits of 3D and 3D + t shape information using SSM with the power of utilizing transfer learning by previously trained CNNs. MethodsIn this study, we employed a dataset provided for the MI classification challenge using SSM in the form of 3D surface points. We proposed to convert these 3D points as different projections per view (PPV) on the x-y, x-z, and y-z planes. Besides, additional 3D and 3D + t dedicated information, i.e., thickness and displacement were added in color on 2D PPVs. We also proposed two augmentation strategies for shape and color based on the SSM concept. Given the optimal combination of the PPVs and colors, our prepared 2D images were fed into the DenseNet-121, DenseNet-169, and DenseNet-201, independently. Four different fine-tuning scenarios (i.e., scenario-I to scenario-IV) were investigated, compared, and optimized to distinguish MI patients from healthy subjects. ResultsThe experimental results evaluated the performance of the different transfer learning scenarios in MI detection. The results demonstrated that the optimized DenseNet-201-Scenario-II achieved the best performance with the accuracy, sensitivity, specificity, precision, F1-score of 0.942 ± 0.039, 0.899 ± 0.060, 1.000 ± 0.000, 1.000 ± 0.000, and 0.946 ± 0.034, respectively in the MI detection task. ConclusionThe promising results prove that combining the use of images with information fusion, the power of the SSM concept in the augmentation process, and the benefits of transfer learning using CNNs, leads to developing a complementary and automatic MI diagnosis tool that can be extended to other diagnostic applications in the future.