OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines.

Objective.To establish an open framework for developing plan optimization models for knowledge-based planning (KBP).Approach.Our framework includes radiotherapy treatment data (i.e. reference plans) for 100 patients with head-and-neck cancer who were treated with intensity-modulated radiotherapy. That data also includes high-quality dose predictions from 19 KBP models that were developed by different research groups using out-of-sample data during the OpenKBP Grand Challenge. The dose predictions were input to four fluence-based dose mimicking models to form 76 unique KBP pipelines that generated 7600 plans (76 pipelines × 100 patients). The predictions and KBP-generated plans were compared to the reference plans via: the dose score, which is the average mean absolute voxel-by-voxel difference in dose; the deviation in dose-volume histogram (DVH) points; and the frequency of clinical planning criteria satisfaction. We also performed a theoretical investigation to justify our dose mimicking models.Main results.The range in rank order correlation of the dose score between predictions and their KBP pipelines was 0.50-0.62, which indicates that the quality of the predictions was generally positively correlated with the quality of the plans. Additionally, compared to the input predictions, the KBP-generated plans performed significantly better (P< 0.05; one-sided Wilcoxon test) on 18 of 23 DVH points. Similarly, each optimization model generated plans that satisfied a higher percentage of criteria than the reference plans, which satisfied 3.5% more criteria than the set of all dose predictions. Lastly, our theoretical investigation demonstrated that the dose mimicking models generated plans that are also optimal for an inverse planning model.Significance.This was the largest international effort to date for evaluating the combination of KBP prediction and optimization models. We found that the best performing models significantly outperformed the reference dose and dose predictions. In the interest of reproducibility, our data and code is freely available.

Convolutional neural network-based dosimetry evaluation of esophageal radiation treatment planning.

PURPOSE: A dosimetry evaluation model for treatment planning of esophageal radiation therapy is developed using a deep learning model. The model predicts dose volume histogram (DVH) from distance to target histogram (DTH) based on stacked de-noise auto-encoder (SDAE) and one-dimensional convolutional network (1D-CN). METHOD: First, SDAE is used to extract the features from the curves of DTH and DVH. Then 1D-CN model is employed to learn the relationship between the features of DTH and DVH, and later used to predict the features of DVH from the features of DTH. Finally, the curve of DVH is restored from the features of DVH based on SDAE. Two hundred and seventy treatment plans are used for training 1D-CN and another sixty-three treatment plans are used for evaluating this model. This method is also compared with another two popular prediction methods based on support vector machine (SVM) and U-net. RESULTS: Based on the experimental result, the proposed model achieves the lowest dose endpoint error comparing to the other models. The average prediction error on planned target volume, left lung, right lung, heart, and spinal cord is 2.94% for the proposed model, while the average prediction errors are 6.79% and 3.41% for SVM and U-net, respectively. CONCLUSIONS: A dosimetry evaluation method based on SDAE and 1D-CN is developed in characterizing the correlation relationship between DTH and DVH of treatment plans. The results show that the model could be trained more efficiently in this framework and the DVH could be predicted with higher accuracy comparing to those existing methods. It provides a useful tool in supporting automated treatment planning of esophageal intensity-modulated radiotherapy.

Deep Learning Based Dosimetry Evaluation at Organs-at-Risk in Esophageal Radiation Treatment Planning.

Rapid esophageal radiation treatment planning is often obstructed by manually adjusting optimization parameters. The adjustment process is commonly guided by the dose-volume histogram (DVH), which evaluates dosimetry at planning target volume (PTV) and organs at risk (OARs). DVH is highly correlated with the geometrical relationship between PTV and OARs, which motivates us to explore deep learning techniques to model such correlation and predict DVHs of different OARs. Distance to target histogram (DTH) is chosen to measure the geometrical relationship between PTV and OARs. DTH and DVH features are then undergone dimension reduction by autoencoder. The reduced feature vectors are finally imported into deep belief network to model the correlation between DTH and DVH. This correlation can be used to predict DVH of the corresponding OAR for new patients. Validation results revealed that the relative dose difference of the predicted and clinical DVHs on four different OARs were less than 3%. These promising results suggested that the predicted DVH could provide near-optimal parameters to significantly reduce the planning time.

An Augmented Cell Segmentation in Fluorescent in Situ Hybridization Images.

Fluorescence in situ hybridization (FISH) surpass previously available technology to become a foremost biological assay, which can provide reliable imaging biomarkers to diagnose cancer and genetic disorders in the cellular level. In order to guarantee the validity of the quality analysis in cell images, it is significant to accurately segment the cell touching regions. We previously structured a mini-U-net to precisely capture cell regions, but this method sometimes can not separate multiple cells that are attached to each other. This work aims to solve this matter by applying cell identification results to provide more accurate prior information for the watershed to describe the cell boundaries. Validation results on 458 cells showed that Dice coefficients and intersection over union were improved from 81.92% to 83.98% and from 68.34% to 73.83% (p=0.03), respectively. The improved results indicated that cell identification is an effective means to handie the cell touching and produce more accurate cell segmentation.