Enhancing automated right-sided early-stage breast cancer treatments via deep learning model adaptation without additional training

BACKGROUND Input data curation and model training are essential, but time-consuming steps in building a deep-learning (DL) auto-planning model, ensuring high-quality data and optimized performance. Ideally, one would prefer a DL model that exhibits the same high-quality performance as a trained model without the necessity of undergoing such time-consuming processes. That goal can be achieved by providing models that have been trained on a given dataset and are capable of being fine-tuned for other ones, requiring no additional training. PURPOSE To streamline the process for producing an automated right-sided breast (RSB) treatment planning technique adapting a DL model originally trained on left-sided breast (LSB) patients via treatment planning system (TPS) specific tools only, thereby eliminating the need for additional training. METHODS The adaptation process involved the production of a predicted dose (PD) for the RSB by swapping from left-to-right the symmetric structures in association with the tuning of the initial LSB model settings for each of the two steps that follow the dose prediction: the predict settings for the post-processing of the PD (ppPD) and the mimic settings for the dose mimicking, respectively. Thirty patients were involved in the adaptation process: Ten manual plans were chosen as ground truth for tuning the LSB model settings, and the adapted RSB model was validated against 20 manual plans. During model tuning, PD, ppPD, and mimicked dose (MD) were iteratively compared to the manual dose according to the new RSB model settings configurations. For RSB model validation, only MD was involved in the planning comparison. Subsequently, the model was applied to 10 clinical patients. Manual and automated plans were compared using a site-specific list of dose-volume requirements. RESULTS PD for the RSB model required substantial corrections as it differed significantly from manual doses in terms of mean dose to the heart (+11.1 Gy) and right lung (+4.4 Gy), and maximum dose to the left lung (+6.4 Gy) and right coronary (+11.5 Gy). Such discrepancies were first addressed by producing a ppPD always superior to the manual dose by changing or introducing new predict settings. Second, the mimic settings were also reformulated to ensure a MD not inferior to the manual dose. The final adapted version of the RSB model settings, for which MD was found to be not significantly different than the manual dose except for a better right lung sparing (-1.1 Gy average dose), was retained for the model validation. In RSB model validation, a few significant-yet not clinically relevant-differences were noted, with the right lung being more spared in auto-plans (-0.6 Gy average dose) and the maximum dose to the left lung being lower in the manual plans (-0.8 Gy). The clinical plans returned dose distributions not significantly different than the validation plans. CONCLUSION The proposed technique adapts a DL model initially trained for LSB cancer for right-sided patients. It involves swapping the dose predictions from left to right and adjusting model settings, without the need for additional training. This technique-specific to a TPS-could be transposed to other TPS platforms.