Exploring Federated Deep Learning for Standardising Naming Conventions in Radiotherapy Data

Abstract: Standardising structure volume names in radiotherapy (RT) data is necessary to enable data mining and analyses, especially across multi-institutional centres. This process is time and resource intensive, which highlights the need for new automated and efficient approaches to handle the task. Several machine learning-based methods have been proposed and evaluated to standardise nomenclature. However, no studies have considered that RT patient records are distributed across multiple data centres. This paper introduces a method that emulates real-world environments to establish standardised nomenclature. This is achieved by integrating decentralised real-time data and federated learning (FL). A multimodal deep artificial neural network was proposed to standardise RT data in federated settings. Three types of possible attributes were extracted from the structures to train the deep learning models: tabular, visual, and volumetric. Simulated experiments were carried out to train the models across several scenarios including multiple data centres, input modalities, and aggregation strategies. The models were compared against models developed with single modalities in federated settings, in addition to models trained in centralised settings. Categorical classification accuracy was calculated on hold-out samples to inform the models performance. Our results highlight the need for fusing multiple modalities when training such models, with better performance reported with tabular-volumetric models. In addition, we report comparable accuracy compared to models built in centralised settings. This demonstrates the suitability of FL for handling the standardization task. Additional ablation analyses showed that the total number of samples in the data centres and the number of data centres highly affects the training process and should be carefully considered when building standardisation models.

• The paper introduces a federated deep learning framework that standardizes RT naming conventions while maintaining data privacy across multiple centers.
• It employs a multimodal approach by combining tabular, 2D, and 3D imaging data to enhance model accuracy comparable to centralized systems.
• The evaluation of FL aggregation methods, including FedAvg and FedAdam, underscores the influence of data distribution on model performance.

Federated Learning Application for Standardizing Radiotherapy Nomenclature

Introduction to Federated Learning in Radiotherapy Data Standardization

The application of federated learning (FL) techniques to standardize naming conventions in radiotherapy (RT) data represents an innovative approach in the field of medical data analysis. The complexity of handling RT patient records, which are distributed across multiple institutions with stringent data privacy regulations, necessitates novel methods that respect these constraints while enabling effective data mining and analysis. This paper introduces a multimodal deep learning model that operates under a federated learning framework to address the challenges of standardizing structure volume names within RT data, a critical step for facilitating data mining and analyses across multi-institutional centers.

Proposal and Methodology

Data Collection and Preprocessing

The study utilized a dataset of lung cancer patients from The Cancer Imaging Archive (TCIA), focusing on standardizing seven specified classes, including one target volume (TV) and six organs-at-risk (OARs). Features extracted included tabular, visual (2D central slices), and volumetric (3D) data from the contoured volumes, which provided a rich representation necessary for the deep learning model.

Model Architecture

The multimodal deep learning model proposed employs a layer-level fusion approach where tabular, visual, and volumetric modalities are concatenated within the neural network. This structure leverages the complementary information provided by each data type, with the convolutional blocks handling imaging data and fully connected layers processing tabular features.

Federated Learning Framework

The federated learning setup involved a centralized orchestrator coordinating model training across simulated data centers, maintaining data privacy by keeping patient records localized. Several FL aggregation strategies were evaluated, including FedAvg, FedOpt, FedYogi, and FedAdam, to determine their effectiveness in this context.

Experimental Findings

Model Performance and Comparison

The findings underscored the necessity of integrating multiple modalities for improved model performance, with tabular-volumetric models notably outperforming other combinations. Even within a federated learning environment, models achieved comparable accuracy to centralized approaches, with a significant classification accuracy evident when employing multimodal inputs. The number of data centers and samples significantly influenced model training, underscoring the importance of strategic data distribution and aggregation method selection in federated settings.

Practical Implications

This research illuminates the feasibility and efficacy of using federated deep learning for the standardization of naming conventions in RT data. It proves that despite the distributed nature of the data, substantial performance can be achieved, comparable to that of centralized models. Furthermore, it suggests the potential of FL in overcoming data privacy and security challenges inherent in multi-institutional healthcare data handling.

Future Directions and Limitations

While the study presents a robust foundation, future research could explore real-world applications involving data from actual distributed centers, encompass broader class representations, and investigate the potential of few-shot learning to handle scenarios with limited labeled data. Additionally, employing augmentation techniques could further enhance model performance and generalizability.

Conclusion

The exploration of federated deep learning for RT data standardization signifies a promising step toward harnessing the power of distributed medical datasets without compromising data privacy. The study’s findings emphasize the viability of FL in medical data analysis, offering a pathway to more personalized and effective cancer treatment planning through standardization of RT data across institutions.