Background: Effective radiation treatment of cancer requires maximising the dose to target while minimising healthy tissue toxicity. For prostate cancer (PC) treatment, the prostate position is confirmed each day using kV imaging to visualise gold seed fiducial markers (FMs) inserted into the prostate before treatment simulation and visible on the CT-based radiation treatment plan. With the increasing utilisation of MRI for radiation therapy planning, a transition to an MRI-only workflow for PC is desirable. However, unambiguous identification of FMs with MRI can be problematic in patients with prostatic calcifications. This work investigates a deep learning-based approach to FM detection using quantitative susceptibility mapping (QSM) alongside conventional MR images.
Method: A 3D U-Net model was built using fastai and MONAI libraries. It processed single or multimodal input images to generate probability maps for background, calcification, and FM regions. The architecture had five levels with channel sizes (16, 32, 64, 128, 256) and strides (2, 2, 2, 2), featuring two residual units per level. Data augmentation techniques included flipping, affine transformations, and normalisation. Input images and masks were cropped to 80x80x80 voxels. The U-Net was trained on individual modalities (CT, QSM, T1, SWI, GRE) and a multimodal model (QSM and T1). The dataset was split into training (12) and validation sets (3) using k-fold cross-validation with k=6. Training used a batch size of 4. The loss function combined Dice and cross-entropy losses with one-hot encoding. Optimisation used RAdam with Lookahead (Ranger), with an initial learning rate of 0.003 and a cosine annealing schedule for 700 epochs. The final model was chosen based on the best validation score. Model performance was evaluated across the cross-validation sets using the Area Under the Receiver Operating Characteristic curve (AUC-ROC) generated at the voxel level for the FM label, and average precision/recall values generated at the FM level.
Results: The voxel-level ROC-AUCs include CT: 1.0, QSM+T1: 0.91, QSM: 0.89, T1: 0.86, SWI: 0.71 and GRE: 0.67. FM-level precision/recall values of interest include CT: 0.98±0.05/1.0±0.0, QSM+T1: 0.76±0.14/0.76±0.08, QSM: 0.77±0.18/0.74±0.13, T1: 0.71±0.19/0.84±0.21.
Discussion: The models trained using QSM images outperform those trained with conventional MRI for the FM segmentation task. The developed models offer a potential solution for MRI-only radiation therapy planning for PC. Future work will focus on increasing the model accuracy by expanding the data training set and developing an image processing pipeline to segment FMs automatically for contouring purposes.