A deep learning-based approach to detect and segment osteolytic bone lesions in whole-body, low-dose CT imaging of multiple myeloma patients

Short Summary: We are developing deep learning architecture which shows promising results in detecting and segmenting osteolytic bone lesion sites in low dose, full body CT scans. By segmenting the lesions, we obtain useful information for medical experts such as lesion size and shape characteristics.
Purpose/Objectives: To develop automated solutions to segment osteolytic lesions and can indicate potential lesions sites within these CT-scans. These lesions are hard to detect and yet very relevant for the diagnosis of mutiple myeloma. Furthermore, we expect that the application of our work to be much broader than this disease. We forsee that our architecture can be applied to detect osteolytic osseous metastases because of the resemblance of these lesions on CT images, and the fact that many oncology patients are also examined and followed with CT scans of neck-chest and abdomen.
Methods and materials: Our dataset consists of 96 low dose CT scans from 79 patients, comprising of 37 full body, 42 upper body and 17 lower body scans. These scans were retrieved from the PACS system of Elisabeth Tweesteden Ziekenhuis (ETZ) in Tilburg, Netherlands and privacy-sensitive metadata were anonymized. The inclusion criteria are patients older than 18 years old, presenting at lesions which are 5 millimetres in diameter or larger. 663 lesions were annotated by one experienced radiologist and 3 radiology residents at ETZ hospital using the segmentation tool in 3D Slicer on the axial slices.
We trained a U-Net model on patches of 192 × 192 pixels on the axial slices. From the slices in which annotated lesions are found, ten random patches are extracted in which a lesion is present, and data augmentation was applied 4 times on each patch to further increase the training dataset size. The utilized data augmentation methods were; horizontal and vertical flipping and rotation, which increased the number of training samples to approximately 110,000 patches. After training, the method was evaluated by calculating the model’s detection rate and Dice scores when applying a sliding window method to complete CT-scans in the test set. To prevent overfitting, the model was trained and evaluated using 5-fold cross-validation with balanced hold out sets.
Results: Averaging the results over the five folds showed that the proposed method managed to detect 89% of the annotated lesions and achieved a Dice score of roughly 0.60.
Conclusion: The method is capable of detecting and segmenting osteolytic bone lesions on various locations within low-dose CT-scans. As the number of patients included in this study is limited, we expect that the result will improve after including new patients in the training data.