Motivation: Public challenge datasets are beneficial to improve classification and segmentation by computer aided diagnosis (CAD); transfer learning may further improve CAD clinical performance. Goal(s): To investigate if transfer learning can improve the accuracy of whole-prostate and lesion segmentation in multi-parametric MRI. Approach: Two nnU-Net networks for whole-prostate and lesion segmentation were initially trained on the PROSTATEx public dataset, then fine-tuned with an in-house clinical dataset. Results: Fine-tuning on clinical data improved the mean Dice score for whole-prostate segmentation from 0.783 to 0.898. However, lesion segmentation networks underperformed due to dataset variability, indicating that while transfer learning is promising, lesion segmentation needs further refinement. Impact: A nnU-Net network trained on large public datasets, then fine-tuned with a small clinical dataset improved whole-prostate segmentation. This network will facilitate processing requiring whole-prostate masks, such as Quantitative Susceptibility Mapping, and could potentially reduce radiological workload or automate quantification.
Ali et al. (Tue,) studied this question.