BronchoPose: an analysis of data and model configuration for vision-based bronchoscopy pose estimation. (January 2023)

• Record Type: Journal Article
• Authors: Borrego-Carazo, Juan; Sanchez, Carles; Castells-Rufas, David; Carrabina, Jordi; Gil, Débora
• Is Part Of: Computer methods and programs in biomedicine. Volume 228(2023)
• Publication Date: 2023-01
• Subjects: Video bronchoscopy guiding -- Deep learning -- Architecture optimization -- Datasets -- Standardized evaluation framework -- Pose estimation; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2022.107241
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: Synthetic Dataset. Publication of a bronchoscopy navigation synthetic dataset based on real anatomies to enable fair comparison among methods. As far as the authors are aware, it is the first public synthetic dataset for bronchoscopy navigation. Evaluation Protocols. A study and comparison of rotation and position losses and metrics (including a novel one) for bronchoscopy navigation, which helps to establish better grounds for training and evaluation. Processing of Temporal Information. Research of different solutions for neural network temporal learning in bronchoscopy navigation (LSTM, Convolutional LSTM, and 3D convolutions) to exploit temporal information in bronchoscopic videos, which is currently under explored. Population Modelling. Analysis of the different options with regards to data usage for industrial applications: first, with a across patient setting, focused on the development of a general model, and secondly, with an intra-patient setting, by providing a specialized model. Abstract: Background and objective : Recent advances in neural networks and temporal image processing have provided new results and opportunities for vision-based bronchoscopy tracking. However, such progress has been hindered by the lack of comparative experimental data conditions. We address the issue by sharing a novel synthetic dataset, which allows for a fair comparison of methods. Moreover, as incorporating deep learning advances in temporal structures is not yet explored in bronchoscopy navigation, we investigate several neural network architectures for learning temporal information at different levels of subject personalization, providing new insights and results. Methods : Using our own shared synthetic dataset for bronchoscopy navigation and tracking, we explore deep learning temporal information architectures (Recurrent Neural Networks and 3D convolutions), which have not been fully explored on bronchoscopy tracking, putting a special focus on network efficiency by using a modern backbone (EfficientNet-B0) and ShuffleNet blocks. Finally, we provide a study of different losses for rotation tracking and population modeling schemes (personalized vs. population) for bronchoscopy tracking. Results : Temporal information architectures provide performance improvements, both in position and angle estimation. Additionally, population scheme analysis illustrates the benefits of offering a personalized model, while loss analysis indicates the benefits of using an adequate metric, improving results. We finally compare with a state-of-the-art model obtaining better results both in performance, with 12.2 % and 18.7 % improvement for position and rotation respectively, and around 67.6 % reduction in memory consumption. Conclusions : Proposed advances in temporal information architectures, loss configuration, and population scheme definition allow for improving the current state of the art in bronchoscopy analysis. Moreover, the publication of the first synthetic dataset allows for further improving bronchoscopy research by enabling proper comparison grounds among methods.