Using dose-surface maps to predict radiation-induced rectal bleeding: a neural network approach

Florian Buettner, S. Gulliford, Steve Webb, Mike PartridgePhysics 2009 · DOI

TL;DR

This paper introduces a novel approach to predict radiation-induced rectal bleeding by using neural networks that analyze the spatial distribution of radiation dose (dose-surface maps or DSMs) to the rectal wall, rather than just volumetric dose metrics. This method shows improved predictive power compared to traditional approaches, suggesting that the "shape" of the dose matters for predicting toxicity.

What Problem Does This Paper Address?

Radiotherapy, while crucial for cancer treatment, can lead to side effects like rectal bleeding. Predicting these toxicities is vital for optimizing treatment plans and patient care. Traditionally, models have relied on dose-volume histograms (DVHs), which summarize the total dose delivered to a certain volume of the organ. However, these DVHs lose information about where within the organ the dose is delivered. This spatial information, or the "shape" of the dose distribution, might be crucial for predicting toxicity. This paper aims to bridge this gap by developing a predictive model that explicitly incorporates the spatial distribution of radiation dose to the rectum.

Methods

The core of this research lies in using dose-surface maps (DSMs) as input for neural networks. DSMs are created by virtually unfolding the rectal wall and projecting the radiation dose onto a 2D surface, capturing the spatial distribution.

Key methodological aspects include:

  • Virtual Rectum-Unfolding: A technique to convert 3D dose distributions into 2D DSMs.
  • Locally-Connected Neural Networks (LCNNs): Unlike fully-connected networks, LCNNs connect nodes in a hidden layer to only a small, local region of the input layer. This is particularly useful for image-like data (like DSMs) as it leverages spatial topology and reduces the number of parameters.
  • Weight Sharing: A technique within LCNNs where the same set of weights is applied to different parts of the input, allowing the network to learn features that are relevant across the entire DSM.
  • Ensemble Learning: Multiple neural networks (250 in this case) are trained, and their predictions are combined (e.g., by voting or averaging) to improve robustness and performance.
  • Expert Ensemble Selection: A subset of the trained networks that demonstrate good generalization ability is selected to form an "expert ensemble," further refining the prediction.
  • Comparison with Dose-Surface Histograms (DSHs): The DSM-based approach is directly compared against a model using DSHs, which are derived from DSMs but represent a more volumetric measure, similar to DVHs.
  • 10-fold Cross-Validation: A standard technique to ensure the model's performance is evaluated on unseen data and to estimate its generalization error.

Datasets

The study utilizes data from the RT01 trial (ISRCTN 47772397), a multicenter randomized controlled trial for prostate cancer patients treated with 3D conformal radiotherapy.

  • Training and Testing Data: Data from 329 patients who were free of rectal bleeding before treatment were used. Of these, 53 patients reported late rectal bleeding post-treatment.
  • Data Splitting: The dataset was split into training and testing sets using a 10-fold cross-validation approach. For ensemble training, a subset of patients from the training set was used.

Key Insights and Ideas

  • The "Shape" of the Dose Matters: The most significant insight is that the spatial distribution (morphology) of the radiation dose to the rectal wall, as captured by DSMs, provides predictive power for rectal bleeding beyond what volumetric measures (like DSHs) offer.
  • LCNNs are Well-Suited for DSMs: Locally-connected neural networks, with their ability to respect spatial topology, are a good fit for analyzing DSMs, potentially extracting more meaningful features than fully-connected networks.
  • Ensemble Learning Enhances Robustness: Combining predictions from multiple neural networks (ensemble learning) leads to more stable and potentially more accurate predictions, especially when dealing with imbalanced datasets or complex relationships.
  • Expert Ensemble Selection Improves Generalization: Selecting a subset of well-generalizing networks from the ensemble can further improve performance and reduce variability.

Key Results

  • DSM-based Classifier Performance: The DSM-based expert ensemble achieved an Area Under the ROC Curve (AUC) of 0.64.
  • DSH-based Classifier Performance: The DSH-based expert ensemble (representing a more volumetric approach) achieved an AUC of 0.59.
  • Performance Improvement: The DSM-based approach demonstrated a noticeable improvement in predictive power (0.64 vs. 0.59 AUC) compared to the DSH-based approach.
  • Variability in Cross-Validation: While the mean AUC was consistent, there was considerable variation in ROC curves between different folds of the cross-validation, attributed to the small number of bleeding events in some test sets and the inherent variance in cross-validation.
  • Expert Selection Benefits: Expert selection resulted in more stable classifiers with less variation in AUC between runs compared to using the full ensemble.

Limitations and Future Work

  • Limitations:
    • Limited Dataset Size: The number of patients who experienced rectal bleeding was relatively small (53 out of 329), which can lead to high variance in performance estimates, especially in cross-validation.
  • Organ Motion Not Considered: The current DSMs are based on planning CT scans and do not account for organ motion during treatment, which can alter the actual dose delivered.
  • Algorithm Dependence: Different algorithms for creating DSMs can result in different maps from the same dose distribution.
  • Output Interpretation: The raw outputs of the neural networks are not direct probabilities and would require rescaling for clinical interpretation.
  • Error Rates: Despite improvements, the error rates are still considered too high for direct clinical application without further refinement.
  • Future Work:
    • Incorporate Organ Motion: Future models could integrate information about organ motion to provide a more accurate representation of the delivered dose.
  • Explore Other DSM Algorithms: Investigating different methods for generating DSMs might yield further improvements.
  • Clinical Translation: Developing methods for rescaling network outputs to provide clinically interpretable probabilities is essential.
  • Larger Datasets: Validation on larger, independent datasets would be crucial to confirm the generalizability of the findings.
  • Non-Dose Factors: While this study focused on dose, incorporating other non-dose factors (age, patient history, etc.) could further enhance predictive accuracy.

Summary updated

Abstract

The incidence of late-toxicities after radiotherapy can be modelled based on the dose delivered to the organ under consideration. Most predictive models reduce the dose distribution to a set of dose-volume parameters and do not take the spatial distribution of the dose into account. The aim of this study was to develop a classifier predicting radiation-induced rectal bleeding using all available information on the dose to the rectal wall. The dose was projected on a two-dimensional dose-surface map (DSM) by virtual rectum-unfolding. These DSMs were used as inputs for a classification method based on locally connected neural networks. In contrast to fully connected conventional neural nets, locally connected nets take the topology of the input into account. In order to train the nets, data from 329 patients from the RT01 trial (ISRCTN 47772397) were split into ten roughly equal parts. By using nine of these parts as a training set and the remaining part as an independent test set, a ten-fold cross-validation was performed. Ensemble learning was used and 250 nets were built from randomly selected patients from the training set. Out of these 250 nets, an ensemble of expert nets was chosen. The performances of the full ensemble and of the expert ensemble were quantified by using receiver-operator-characteristic (ROC) curves. In order to quantify the predictive power of the shape, ensembles of fully connected conventional neural nets based on dose-surface histograms (DSHs) were generated and their performances were quantified. The expert ensembles performed better than or equally as well as the full ensembles. The area under the ROC curve for the DSM-based expert ensemble was 0.64. The area under the ROC curve for the DSH-based expert ensemble equalled 0.59. This difference in performance indicates that not only volumetric, but also morphological aspects of the dose distribution are correlated to rectal bleeding after radiotherapy. Thus, the shape of the dose distribution should be taken into account when a predictive model for radiation-induced rectal bleeding is developed.
BibTeX
@inproceedings{buettner2009using,
  author = {Florian Buettner and S. Gulliford and Steve Webb and Mike Partridge},
  title = {{Using dose-surface maps to predict radiation-induced rectal bleeding: a neural network approach}},
  booktitle = {Physics},
  year = {2009},
  doi = {10.1088/0031-9155/54/17/005},
}
Figure