Moving towards clinical implementation of a deep-learning brain tumour segmentation algorithm: DICOM server integration

Abstract

Aim: This research project is part of a larger initiative where the aim is to develop a software tool that supports radiologists by providing more accurate tumour volume measurements on magnetic resonance imaging (MRI) scans. A specific aim within this larger endeavour is to integrate a deep-learning algorithm for brain tumour segmentation with a digital imaging and communications in medicine (DICOM) server. Method: The DeepMedic segmentation model was trained and tested using the brain tumour segmentation (BraTS) 2024 dataset, which included post-treatment glioma images from four MRI sequences. The model was trained to accurately identify distinct tumour regions. Orthanc was configured and used as a DICOM server to handle medical image retrieval, storage, and reintegration. A Python-based pipeline, built using existing Python packages and libraries, was developed to enable communication between DeepMedic and Orthanc, facilitating automated segmentation and data management. Result: The model was successfully trained on the BraTS 2024 dataset, without overfitting. Testing with the saved model demonstrated its ability to accurately segment the various tumour regions, as evaluated using the Dice similarity coefficient and the 95th percentile of the Hausdorff distance (HD95). While the model performed well in most cases, some discrepancies between regions were observed. An automated workflow integrates DeepMedic with Orthanc. The user enters a study id, after which Orthanc retrieves the corresponding DICOM series, converts, and prepares the images for DeepMedic. After segmentation, the results are converted back to DICOM format and sent to Orthanc for storage and availability for further use. Conclusion: The project demonstrated the potential of the technical implementation of a deep-learning-based brain tumour segmentation algorithm. However, further work is needed to enhance the robustness of the pipeline and ensure its applicability across a wider range of clinical settings before it can be tested on in-house clinical data.