Background

Delivery of the optimised radiation to the patient with precision remained a great problem for decades. Developed an optimised radiation adoption solution for non-small cell lung cancer during clinical decisions treatment plans using deep learning and reinforcement learning AI techniques.

Method

With a retrospective population of 230 NSLC patients who received radiotherapy a 3 component deep learning framework powered by deep reinforcement learning (DRL) for dose fraction escalation and de-escalation is developed. Patient characteristics 9 such as age, weight, stage were considered for radiation dosage decisions. First, TGAN a tabular data synthesizer was used for 10000 patients data generation.  Second, a radiotherapy artificial environment (RAE) was reconstructed by a deep neural network (DNN) utilising both original and synthetic data (by GAN) to predict the dosage values for the patients’ treatment courses. Third, a deep Q -network (DQN) was applied to the RAE for choosing the optimal dose. The rewards were given accordingly w.r.t clinical outcome of better survival in terms of months.

Results

Formed a concordance matrix showcasing the optimised radiation values for the patients for 100 patients The outcome w.r.t survival was used as a baseline to adopt the radiation.

The DQN component freely controlled the estimated adaptive dose per fraction (ranging from 1 ~ 5 Gy). The DRL automatically favoured dose escalation/de-escalation between 1.5 ~ 3.8 Gy, a range similar to that used in the clinical protocol. The results were similar to the radiation protocols depicting that the system is adhering to the protocols and is safe. The reward function based on the survival outcome has made precise and optimised variation in terms of escalation and de-escalation. By adjusting the P+ reward function with higher emphasis on survival, better matching of doses between the DQN and the clinical protocol was achieved with a RMSE = 0.6 Gy. The decisions selected by the DQN seemed to have better concordance with patients’ eventual outcomes compared to the traditional techniques without Q learning.

Conclusion

Optimised & automated dose adaptation system by DRL is practical  and protocol adhered, achieving similar results to those chosen by clinicians. 

Impact statement

Optimised radiation delivery for better quality care with accuracy.