Multimodal radiotherapy dose prediction using a multi-task deep learning model

Background: In radiation therapy (RT), accelerated partial breast irradiation (APBI) has emerged as an increasingly preferred treatment modality over conventional whole breast irradiation due to its targeted dose delivery and shorter course of treatment. APBI can be delivered through various modalities including Cobalt-60-based systems and linear accelerators with C-arm, O-ring, or robotic arm design. Each modality possesses distinct features, such as beam energy or the degrees of freedom in treatment planning, which influence their respective dose distributions. These modality-specific considerations emphasize the need for a quantitative approach in determining the optimal dose delivery modality on a patient-specific basis. However, manually generating treatment plans for each modality across every patient is time-consuming and clinically impractical.

Purpose: We aim to develop an efficient and personalized approach for determining the optimal RT modality for APBI by training predictive models using two different deep learning-based convolutional neural networks. The baseline network performs a single-task (ST), predicting dose for a single modality. Our proposed multi-task (MT) network, which is capable of leveraging shared information among different tasks, can concurrently predict dose distributions for various RT modalities. Utilizing patient-specific input data, such as a patient's computed tomography (CT) scan and treatment protocol dosimetric goals, the MT model predicts patient-specific dose distributions across all trained modalities. These dose distributions provide patients and clinicians quantitative insights, facilitating informed and personalized modality comparison prior to treatment planning.

Methods: The dataset, comprising 28 APBI patients and their 92 treatment plans, was partitioned into training, validation, and test subsets. Eight patients were dedicated to the test subset, leaving 68 treatment plans across 20 patients to divide between the training and validation subsets. ST models were trained for each modality, and one MT model was trained to predict doses for all modalities simultaneously. Model performance was evaluated across the test dataset in terms of Mean Absolute Percent Error (MAPE). We conducted statistical analysis of model performance using the two-tailed Wilcoxon signed-rank test.

Results: Training times for five ST models ranged from 255 to 430 min per modality, totaling 1925 min, while the MT model required 2384 min. MT model prediction required an average of 1.82 s per patient, compared to ST model predictions at 0.93 s per modality. The MT model yielded MAPE of 1.1033 ± 0.3627% as opposed to the collective MAPE of 1.2386 ± 0.3872% from ST models, and the differences were statistically significant (p = 0.0003, 95% confidence interval = [-0.0865, -0.0712]).

Conclusion: Our study highlights the potential benefits of a MT learning framework in predicting RT dose distributions across various modalities without notable compromises. This MT architecture approach offers several advantages, such as flexibility, scalability, and streamlined model management, making it an appealing solution for clinical deployment. With such a MT model, patients can make more informed treatment decisions, physicians gain more quantitative insight for pre-treatment decision-making, and clinics can better optimize resource allocation. With our proposed goal array and MT framework, we aim to expand this work to a site-agnostic dose prediction model, enhancing its generalizability and applicability.