Training deep‐learning segmentation models from severely limited data

Abstract

To enable generation of high-quality deep learning segmentation models from severely limited contoured cases (e.g.,~10 cases). Methods: Thirty head and neck computed tomography (CT) scans with well-defined contours were deformably registered to 200 CT scans of the same anatomic site without contours. Acquired deformation vector fields were used to train a principal component analysis (PCA) model for each of the 30 contoured CT scans by capturing the mean deformation and most prominent variations. Each PCA model can produce an infinite number of synthetic CT scans and corresponding contours by applying random deformations. We used 300, 600, 1000, and 2000 synthetic CT scans and contours generated from one PCA model to train V-Net, a 3D convolutional neural network architecture, to segment parotid and submandibular glands. We repeated the training using same numbers of training cases generated from 7, 10, 20, and 30 PCA models, with the data distributed evenly between each PCA model. Performance of the segmentation models was evaluated with Dice similarity coefficients between auto-generated contours and physician-drawn contours on 162 test CT scans for parotid glands and another 21 test CT scans for submandibular glands.Results: Dice values varied with the number of synthetic CT scans and the number of PCA models used to train the network. By using 2000 synthetic CT scans generated from 10 PCA models, we achieved Dice values of 82.8% AE 6.8% for right parotid, 82.0% AE 6.9% for left parotid, and 74.2% AE 6.8% for submandibular glands. These results are comparable with those obtained from state-of-the-art auto-contouring approaches, including a deep learning network trained from more than 1000 contoured patients and a multi-atlas algorithm from 12 well-contoured atlases. Improvement was marginal when >10 PCA models or >2000 synthetic CT scans were used. Conclusions: We demonstrated an effective data augmentation approach to train high-quality deep learning segmentation models from a limited number of well-contoured patient cases.

Feasibility of Continual Deep Learning-Based Segmentation for Personalized Adaptive Radiation Therapy in Head and Neck Area

“…High-quality training set rather than low-quality but large-volume training set emerges as a simple but effective approach for improving the performance of DLS. Zhao et al proposed synthetic CT generation for training DLS from extremely limited training set [50]. They generated up to 2000 synthetic CT from 30 well-defined segmentations for training DLS resulting in DSC of 0.74-0.83 [50].…”

“…Zhao et al proposed synthetic CT generation for training DLS from extremely limited training set [50]. They generated up to 2000 synthetic CT from 30 well-defined segmentations for training DLS resulting in DSC of 0.74-0.83 [50]. Currently, various DIR software is recommended for ART; DLS is considered as a potential next step in near future [1,51].…”

Knowledge‐based planning for the radiation therapy treatment plan quality assurance for patients with head and neck cancer

“…In addition to the clinical plan used in treatment, an automated plan was generated for each patient using a fully automated TPS, the RadiationPlanning Assistant (RPA). 26 , 27 , 28 , 29 , 30 The RPA is able to generate clinically acceptable VMAT plans with high consistency for patients with head and neck cancer; additional details of the planning process have been published. 31 …”

Intentional deep overfit learning (IDOL): A novel deep learning strategy for adaptive radiation therapy

“…Traditionally, investigators have framed the challenge as combatting overfitting—a scenario in which the model tends to “memorize” the training dataset but performs poorly when applied to unseen data not represented in the training sample. In an effort to prevent this, techniques, such as dropout, batch normalization, data augmentation, and transfer learning, have been widely utilized 5–9 . However, these techniques often fall short when the expressive power of the network is high, when the set of transformations (e.g., translation, rotation) utilized is limited, and when the entirety of the available prior knowledge is not leveraged.…”

“…In an effort to prevent this, techniques, such as dropout, batch normalization, data augmentation, and transfer learning, have been widely utilized. [5][6][7][8][9] However, these techniques often fall short when the expressive power of the network is high, when the set of transformations (e.g., translation, rotation) utilized is limited, and when the entirety of the available prior knowledge is not leveraged.…”