PURPOSE: Two-dimensional radiotherapy is often used to treat cervical cancer in low- and middle-income countries, but treatment planning can be challenging and time-consuming. Neural networks offer the potential to greatly decrease planning time through automation, but the impact of the wide range of hyperparameters to be set during training on model accuracy has not been exhaustively investigated. In the current study, we evaluated the effect of several convolutional neural network architectures and hyperparameters on 2D radiotherapy treatment field delineation. METHODS: Six commonly used deep learning architectures were trained to delineate four-field box apertures on digitally reconstructed radiographs for cervical cancer radiotherapy. A comprehensive search of optimal hyperparameters for all models was conducted by varying the initial learning rate, image normalization methods, and (when appropriate) convolutional kernel size, the number of learnable parameters via network depth and the number of feature maps per convolution, and nonlinear activation functions. This yielded over 1700 unique models, which were all trained until performance converged and then tested on a separate dataset. RESULTS: Of all hyperparameters, the choice of initial learning rate was most consistently significant for improved performance on the test set, with all top-performing models using learning rates of 0.0001. The optimal image normalization was not consistent across architectures. High overlap (mean Dice similarity coefficient = 0.98) and surface distance agreement (mean surface distance < 2 mm) were achieved between the treatment field apertures for all architectures using the identified best hyperparameters. Overlap Dice similarity coefficient (DSC) and distance metrics (mean surface distance and Hausdorff distance) indicated that DeepLabv3+ and D-LinkNet architectures were least sensitive to initial hyperparameter selection. CONCLUSION: DeepLabv3+ and D-LinkNet are most robust to initial hyperparameter selection. Learning rate, nonlinear activation function, and kernel size are also important hyperparameters for improving performance.

Background

Clinical data collection related to prostate cancer (PCa) care is often unstructured or heterogeneous among providers, resulting in a high risk for ambiguity in its meaning when sharing or analyzing data. Ontologies, which are shareable formal (i.e., computable) representations of knowledge, can address these challenges by enabling machine-readable semantic interoperability. The purpose of this study was to identify PCa-specific key data elements (KDEs) for standardization in clinic and research.

Methods

A modified Delphi method using iterative online surveys was performed to report a consensus agreement on KDEs by a multidisciplinary panel of 39 PCa specialists. Data elements were divided into three themes in PCa and included (1) treatment-related toxicities (TRT), (2) patient-reported outcome measures (PROM), and (3) disease control metrics (DCM).

Results

The panel reached consensus on a thirty-item, two-tiered list of KDEs focusing mainly on urinary and rectal symptoms. The Expanded Prostate Cancer Index Composite (EPIC-26) questionnaire was considered most robust for PROM multi-domain monitoring, and granular KDEs were defined for DCM.

Conclusions

This expert consensus on PCa-specific KDEs has served as a foundation for a professional society-endorsed, publicly available operational ontology developed by the American Association of Physicists in Medicine (AAPM) Big Data Sub Committee (BDSC).

Purpose

CT ventilation imaging (CTVI) is being used to achieve functional avoidance lung cancer radiation therapy in three clinical trials (NCT02528942, NCT02308709, NCT02843568). To address the need for common CTVI validation tools, we have built the Ventilation And Medical Pulmonary Image Registration Evaluation (VAMPIRE) Dataset, and present the results of the first VAMPIRE Challenge to compare relative ventilation distributions between different CTVI algorithms and other established ventilation imaging modalities.

Methods

The VAMPIRE Dataset includes 50 pairs of 4DCT scans and corresponding clinical or experimental ventilation scans, referred to as reference ventilation images (RefVIs). The dataset includes 25 humans imaged with Galligas 4DPET/CT, 21 humans imaged with DTPA-SPECT, and 4 sheep imaged with Xenon-CT. For the VAMPIRE Challenge, 16 subjects were allocated to a training group (with RefVI provided) and 34 subjects were allocated to a validation group (with RefVI blinded). Seven research groups downloaded the Challenge dataset and uploaded CTVIs based on deformable image registration (DIR) between the 4DCT inhale/exhale phases. Participants used DIR methods broadly classified into B-splines, Free-form, Diffeomorphisms, or Biomechanical modeling, with CT ventilation metrics based on the DIR evaluation of volume change, Hounsfield Unit change, or various hybrid approaches. All CTVIs were evaluated against the corresponding RefVI using the voxel-wise Spearman coefficient rS , and Dice similarity coefficients evaluated for low function lung ( DSClow ) and high function lung ( DSChigh ).

Results

A total of 37 unique combinations of DIR method and CT ventilation metric were either submitted by participants directly or derived from participant-submitted DIR motion fields using the in-house software, VESPIR. The rS and DSC results reveal a high degree of inter-algorithm and intersubject variability among the validation subjects, with algorithm rankings changing by up to ten positions depending on the choice of evaluation metric. The algorithm with the highest overall cross-modality correlations used a biomechanical model-based DIR with a hybrid ventilation metric, achieving a median (range) of 0.49 (0.27-0.73) for rS , 0.52 (0.36-0.67) for DSClow , and 0.45 (0.28-0.62) for DSChigh . All other algorithms exhibited at least one negative rS value, and/or one DSC value less than 0.5.

Conclusions

The VAMPIRE Challenge results demonstrate that the cross-modality correlation between CTVIs and the RefVIs varies not only with the choice of CTVI algorithm but also with the choice of RefVI modality, imaging subject, and the evaluation metric used to compare relative ventilation distributions. This variability may arise from the fact that each of the different CTVI algorithms and RefVI modalities provides a distinct physiologic measurement. Ultimately this variability, coupled with the lack of a "gold standard," highlights the ongoing importance of further validation studies before CTVI can be widely translated from academic centers to the clinic. It is hoped that the information gleaned from the VAMPIRE Challenge can help inform future validation efforts.