Single patient learning for adaptive radiotherapy dose prediction

Abstract

Background: Throughout a patient's course of radiation therapy, maintaining accuracy of their initial treatment plan over time is challenging due to anatomical changes-for example, stemming from patient weight loss or tumor shrinkage. Online adaptation of their RT plan to these changes is crucial, but hindered by manual and time-consuming processes. While deep learning (DL) based solutions have shown promise in streamlining adaptive radiation therapy (ART) workflows, they often require large and extensive datasets to train population-based models.

Purpose: This study extends our prior research by introducing a minimalist approach to patient-specific adaptive dose prediction. In contrast to our prior method, which involved fine-tuning a pre-trained population model, this new method trains a model from scratch using only a patient's initial treatment data. This patient-specific dose predictor aims to enhance clinical accessibility, thereby empowering physicians and treatment planners to make more informed, quantitative decisions in ART. We hypothesize that patient-specific DL models will provide more accurate adaptive dose predictions for their respective patients compared to a population-based DL model.

Methods: We selected 33 patients to train an adaptive population-based (AP) model. Ten additional patients were selected, and their respective initial RT data served as single samples for training patient-specific (PS) models. These 10 patients contained an additional 26 ART plans that were withheld as the test dataset to evaluate AP versus PS model dose prediction performance. We assessed model performance using Mean Absolute Percent Error (MAPE) by comparing predicted doses to the originally delivered ground truth doses. We used the Wilcoxon signed-rank test to determine statistically significant differences in terms of MAPE between the AP and PS model results across the test dataset. Furthermore, we calculated differences between predicted and ground truth mean doses for segmented structures and determined statistical significance in the differences for each of them.

Results: The average MAPE across AP and PS model dose predictions was 5.759% and 4.069%, respectively. The Wilcoxon signed-rank test yielded two-tailed p-value = $2.9802\times {10}^{-8}$ , indicating that the MAPE differences between the AP and PS model dose predictions are statistically significant, and 95% confidence interval = [-2.1610, -1.0130], indicating 95% confidence that the MAPE difference between the AP and PS models for a population lies in this range. Out of 24 total segmented structures, the comparison of mean dose differences for 12 structures indicated statistical significance with two-tailed p-values < 0.05.

Conclusion: Our study demonstrates the potential of patient-specific deep learning models in application to ART. Notably, our method streamlines the training process by minimizing the size of the required training dataset, as only a single patient's initial treatment data is required. External institutions considering the implementation of such a technology could package such a model so that it only requires the upload of a reference treatment plan for model training and deployment. Our single patient learning strategy demonstrates promise in ART due to its minimal dataset requirement and its utility in personalization of cancer treatment.

Keywords: adaptive; artificial intelligence; deep learning; dose prediction; head and neck cancer; patient; radiation therapy; single.

Figure 1.

Distribution of head and neck patient data for model training in this study.

Figure 2.

Evaluation of model performance for the Adaptive Population (AP) and Patient-specific (PS) models, in terms of Mean Absolute Percent Error (MAPE). Reported MAPE values are an average across the test dataset’s 26 adaptive sessions.

Figure 3.

Dose displays for a single patient’s adaptive session. The greatest prescription dose level is 70Gy for this patient. The top panel displays the ground truth, the middle panel displays the adaptive population (AP) model prediction, and the bottom panel displays the patient-specific (PS) model prediction. Visually, the PS model predicts lower dose than the AP model in critical organs at risk such as the left parotid, right parotid, and superior portions of the nasopharynx and brain. The PS model’s dose prediction appears to closely replicate the intended distribution as observable in the ground truth, which could be a result of patient-specific considerations that the physician had in mind for the sparing of their organs at risk.

Figure 4.

Dose displays for a single patient’s adaptive session. Dose differences are measured in percent difference relative to the greatest prescription dose level – the greatest prescription dose level being 70Gy in this case. The top panel displays the ground truth, and the middle/bottom panels display the dose differences between a given model’s dose prediction and the ground truth dose. Notably, the adaptive population (AP) model dose prediction differences demonstrates overestimation of dose in the left parotid, right parotid, nasopharynx, brain and spinal cord. If the AP model dose prediction was used as guidance in adaptive treatment planning optimization, the patient may receive more dose than necessary to their healthy tissues.

Figure 5.

A sample dose volume histogram (DVH) from the patient as displayed previously in figures #3 and #4. The solid line denotes the ground truth dose, dotted line denotes the adaptive population (AP) model predicted dose, and dashed line denotes the patient-specific (PS) model predicted dose. Large volumetric dose differences can be observed in organs such as the left parotid, right parotid, left submandibular gland and spinal cord. In such organs, the PS model’s dose prediction more closely aligned with the ground truth than the AP model’s dose prediction. By using a PS model for adaptive treatment planning guidance in lieu of an AP model, dose delivered to a patient’s organs at risk may be minimized.

Figure 6.

An example of dose difference displays, subtraction between: 1) a patient’s initial RT (iRT) dose casted onto a patient’s adaptive session scan, and 2) the PS model’s predicted dose for that patient’s adaptive session. This patient has three segmented planning target volumes (PTVs), which are displayed in the axial view (subplot A), coronal view (subplot C), and sagittal view (subplot D). The PTV dose volume histograms (DVH) for the casted iRT dose and the PS model’s predicted dose are visible in subplot B. These subplots serve as visual aid to demonstrate the uniqueness of the PS model’s predicted dose as compared to the casted iRT dose.