| dc.description.abstract | Purpose: Lung cancer, both primary and secondary, has one of the highest cancer-related mortality rates. This form of cancer can be detected using the high-resolution medical imaging technique, Computed Tomography (CT). Sometimes, even for health professionals, these lung lesions might be difficult to spot, thus providing segmentations, which may lead to late lesion detection and death. This is where deep learning (DL) can provide great assistance and prevent unnecessary deaths. In this thesis, the purpose was, therefore, to train and evaluate the DL algorithm You Only Look Once (YOLO) for CT lung lesion segmentation (and detection) and establish a local clinical workflow utilizing the algorithm.
Methods: This thesis was conducted by adapting the large online LIDC-IDRI dataset containing CT images of lung cancer patients to fit the utilized DL approach. After the images were pre-processed, models with various complexities were trained on datasets both with and without lesion size restrictions. Finally, a trained model was uploaded to the regional clinical research PACS, and the model was applied to local test data.
Results: Some of the trained models managed to detect and segment lung lesions to a high degree. The models trained with size restrictions tended to perform better. Additionally, the model uploaded to the regional research PACS was tested on unlabeled CT lung test images and was able to segment (and detect) what seemed to be tumors.
Conclusion: The trained model uploaded demonstrated the ability to detect and segment lung lesions, both on the online data and the test data, from the regional research PACS system, suggesting further exploring and optimizing the model developing methodology for future work. Additionally, after optimizing the model one could try to train the model to predict tumor development longitudinally. | |
