Radiotherapy dosage: A neural network approach for uninvolved liver dose in stereotactic body radiation therapy for liver cancer

Abstract

A recent study by Zhang et al developed a neural network-based predictive model for estimating doses to the uninvolved liver during stereotactic body radiation therapy (SBRT) in liver cancer. The study reported a significant advancement in personalized radiotherapy by improving accuracy and reducing treatment-related toxicity. The model demonstrated strong predictive performance with R-values above 0.8, indicating its potential to improve treatment consistency. However, concerns arise from the small sample size and exclusion criteria, which may limit generalizability. Future studies should incorporate larger, more diverse patient cohorts, explore potential confounding factors such as tumor characteristics and delivery technique variability, and address the long-term effects of SBRT.

Key Words: Liver cancer; Radiotherapy dosage; Dose prediction; Machine learning; Stereotactic body radiotherapy

Core Tip: A study by Zhang et al developed a neural network-based predictive model for estimating doses to uninvolved liver tissue during stereotactic body radiation therapy (RT), representing a significant advancement in personalizing RT for liver cancer patients. The model demonstrated high predictive accuracy, with R-values exceeding 0.8, highlighting its potential to standardize dose estimation and improve patient safety by reducing biases. The study's relatively small patient cohort (114 patients) raises concerns about selection bias and limits the model's generalizability. Future research should involve larger multicenter cohorts and a more comprehensive cohort of patient characteristics to improve the generalizability of models and clinical relevance. Interdisciplinary collaboration among oncologists, data scientists, and radiation technologists is vital for improving predictive models and the efficacy and precision of cancer treatment.

TO THE EDITOR

We read the study by Zhang et al[1] with great interest and congratulate the authors on their work: Liver cancer ranks as the second most common cause of premature death from cancer[2]. The treatment modalities include radiation therapy (RT), radiofrequency ablation, transarterial chemoembolization, transarterial radioembolization, targeted therapy, immunotherapy, hepatectomy, and liver transplant[2]. The role of RT has been refined with the increasing use of stereotactic body RT (SBRT)[3]. SBRT focuses high energy on the target with better sparing of the surrounding tissue[4]. The authors discussed using a neural network model for predicting doses to the uninvolved liver during SBRT in liver cancer; this innovative approach represents a significant advancement in personalized cancer treatment. While acknowledging the authors' rigorous efforts and valuable contributions as the study addresses reducing treatment-related toxicity while optimizing therapeutic efficacy, we offer constructive suggestions for further refinement.

The study offered a notable advancement, introducing a neural network-based predictive model for estimating doses to uninvolved liver tissue during SBRT. Using objective data rather than relying solely on the planner's subjective experience helps reduce biases affecting treatment quality. As reported, the model shows a strong correlation (R-values above 0.8) between predicted and actual doses, indicating its potential to standardize treatment outcomes and improve patient safety in clinical practice[1]. Moreover, using stratified random sampling to maintain consistency in patient characteristics across training, validation, and test datasets improves the model's reliability, laying a strong foundation for future studies to minimize variations in treatment outcomes.

Despite these advancements, the manuscript highlights significant concerns regarding potential bias and confounding factors. The study included a relatively small cohort of 114 patients, which raises concerns about selection bias. Furthermore, the small sample size increases the risk of overfitting the model and limiting its generalizability. In addition, inclusion criteria may limit the applicability of findings to a broader patient population. For instance, excluding patients with concurrent malignancies or those experiencing motion artifacts could result in omitting important data that might impact the strength of the predictive model. Future studies should aim for a larger, more diverse patient population, potentially through multicenter collaborations, to improve the generalizability of the findings. Similarly, extended follow-up studies can provide valuable information about the long-term effects of treatment, potential late toxicities, and overall survival outcomes.

In addition, more comprehensive potential confounding factors should be included. Patient characteristics such as tumor size, location, and anatomical differences were not fully accounted for, which may confound the relationship between the predicted and actual doses received by the uninvolved liver. Future research should incorporate a comprehensive analysis of clinical variables and consider using stratified statistical analyses to understand their impact on treatment outcomes[5]. We believe that integrating artificial intelligence with RT planning holds important promise. As machine learning algorithms advance, they may provide sophisticated tools for predicting treatment outcomes, thus improving the planning process[6]. Collaborative efforts among oncologists, data scientists, and radiotherapy technologists are essential in developing these next-generation predictive models[7]. Additionally, investigating the ethical implications of automated decision-making in patient care remains vital, confirming that patient safety and autonomy should be maintained.

In the context of future research, variations in the radiation delivery techniques among different practitioners could introduce confounding factors affecting the treatment outcomes[8]. The study's design did not explore this aspect in detail. Hence, it is essential to standardize treatment delivery protocols and train practitioners to reduce differences arising from technical variability.

CONCLUSION

In conclusion, the study by Zhang et al[1] demonstrated a significant advancement in radiation therapy planning; we urge researchers and practitioners to remain cautious regarding potential biases and confounding factors, which are important for improving the reliability and applicability of their findings. The suggestions provided aim to address these challenges; future studies can refine predictive models and contribute to improved patient outcomes. By addressing these challenges head-on, the field can continue to progress toward more personalized and effective cancer treatment approaches. Future studies should develop adaptive models that can continually learn from new data, improving their predictive abilities as more treatment outcomes are recorded.