gms | German Medical Science

Introduction: The urgency for early detection of liver cancer is paramount, significantly increasing survival rates. Recent advances in artificial intelligence (AI) have proposed more effective early detection methods. By analysing medical imaging scans, such as CT scans, AI algorithms can identify abnormalities indicative of cancerous growths with remarkable accuracy. However, the efficacy of these algorithms heavily relies on the availability of labelled data and a massive amount of this available data for model training purposes.

Moreover, relevant data for model training is distributed in multiple sources. This work presents an implementation of federated liver and tumour segmentation models using Personal Health Train (PHT) infrastructure.

Federated Model using PHT Infrastructure: Our federated models consist of training a model based on distributed data. PHT [1], [2] infrastructure provides access to different endpoints and makes distributed operations (including federated learning).

We define three PHT Stations (clients) at different places (Leipzig, Cologne, and Aachen) and a central service (central server). Each PHT Station is an independent machine responsible for training a local model based on its locally stored data. Subsequently, the PHT Central Service is connected to each Station and aggregates the trained local models, proposing a global model through local models' weight averaging.

Data Setup: We started training a segmentation model based on the medical segmentation decathlon dataset [3] stored in a centralised server. It contains 131 3D-labelled NIFTI CT images converted into thousands of valid 2D scans and split into three equal parts (each part stored in a specific PHT Station). Each image contains segmentation masks that help identify the regions where the liver and tumours are. We also implemented some state-of-the-art segmentation models (UNET [4], DenseUNET [5], ResUNET++ [6], and Attention-Unet [7]).

We separated the CT scans into training (80%) and test (20%) sets for each PHT Station. We used the dice for liver and tumour regions and the IoU measures to evaluate each model's performance.

Results: We trained and evaluated the previously mentioned segmentation models on the specified dataset. Remarkably, all three segmentation models demonstrated comparable performance. Every model is active more than 90% of the time for dice score prediction of the liver region and more than 70% for dice score prediction of tumour regions for different endpoints.

Discussion: Predictive models such as image segmentation, particularly in sensitive medical domains such as liver cancer segmentation, demand high efficiency in real-world applications. Performance metrics such as the dice coefficient must exceed 90% for liver and tumour region prediction on new images to be adopted in medical institutions to help doctors in decision-making. Incorporating additional CT scans becomes imperative to enhance segmentation model robustness and generalisation. However, amalgamating diverse datasets into a centralised server poses significant challenges, primarily due to stringent data privacy regulations across different institutions, such as DIC Leipzig and Cologne. In the future, we want to join other data sources (e.g., hospitals) containing CT Liver images in our experiments to enrich our federated model.

Acknowledgement: This work is supported by the BMBF project FAIR Data Spaces (FAIRDS14).

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.