Training Convolutional Networks for Prostate Segmentation With Limited Data

Abstract

Multi-zonal segmentation is a critical component of computer-aided diagnostic systems for detecting and staging prostate cancer. Previously, convolutional neural networks such as the U-Net have been used to produce fully automatic multizonal prostate segmentation on magnetic resonance images (MRIs) with performance comparable to human experts, but these often require large amounts of manually segmented training data to produce acceptable results. Thus for institutions that have limited amounts of labeled MRI exams, it is not clear how much data is needed to train a segmentation model, and which training strategy should be used to maximize the value of the available data. This work compares how the strategies of transfer learning and aggregated training using publicly available external data can improve segmentation performance on internal, site-specific prostate MR images, and evaluates how the performance varies with the amount of internal data used for training.Cross training experiments were performed to show that differences between internal and external data were impactful. Using a standard U-Net architecture, optimizations were performed to select between 2D and 3D variants, and to determine the depth of fine-tuning required for optimal transfer learning. With the optimized architecture, the performance of transfer learning and aggregated training were compared for a range of 5-40 internal datasets. The results show that both strategies consistently improve performance and produced segmentation results that are comparable to that of human experts with approximately 20 site-specific MRI datasets. These findings can help guide the development of site-specific prostate segmentation models for both clinical and research applications.

A bi-directional segmentation method for prostate ultrasound images under semantic constraints

“…This creates a complicated issue when it comes to existential gradients in label-based deep learning models. Currently, numerous proprietary datasets are available for prostate segmentation 14,34,35 , Our approach avoids the above drawbacks and utilizes the feature information of the image itself for segmentation. Compared to traditional deep learning models, the method is not limited by the size of the dataset and learns error information without the influence of annotations.…”

A systematic review of deep learning methods for the classification and segmentation of prostate cancer on magnetic resonance images

“…Utilizing datasets of varying sizes and characteristics, including both public datasets and private datasets, these papers emphasize achieving high diagnostic accuracy in PCa detection tasks. Further, papers 34–64 focus on segmentation accuracy, evaluating methods using metrics like DSC, Hausdorff Distance (HD), or Intersection over Union (IoU). Additionally, they compare segmentation results with existing methods, highlighting improvements in accuracy, efficiency, or robustness, while emphasizing clinical implications for treatment planning and patient outcomes.…”

“…Saunders et al 62 compared aggregated training and transfer learning to enhance segmentation accuracy for site‐specific prostate MRIs with limited labeled data, utilizing an optimized U‐Net architecture. Both strategies consistently demonstrated strong performance, resembling human expert segmentation outcomes, promising for PCa detection and segmentation.…”

Machine Learning in Prostate MRI for Prostate Cancer: Current Status and Future Opportunities

“…ENet (0.91, 0.87, 0.71) and UNet (0.88, 0.86, 0.70) were more accurate than ERFNet (0.87, 0.84, 0.65) in terms of DSC (for WG, TZ and PZ, respectively), while ENet outstood the other two methods, with faster convergence speed and fewer parameters. Saunders et al [ 49 ] compared the performance of independent training, transfer learning, and aggregated learning based on 3D and 2D U-Net models, on the premise of limited training data. In addition, 3D U-Net was found to be more robust to a small sample size (five training cases) than 2D U-Net by an average DSC of 0.18, while transfer learning and aggregated learning (similar DSC: 0.73, 0.83, 0.88 for PZ, CG, WG, respectively) both outperformed independent training (DSC 0.65, 0.77, 0.83) when using five internal training cases.…”