Modeling focused kV X-ray dose kernels via 3D implicit neural functions for small animal brain radiation

Precise irradiation is critical in radiation-based neuromodulation and neuro-oncology for small animals like mice. The focused kV x-ray approach was proposed to achieve a restricted target size that is difficult for existing devices. The treatment planning of focused X-ray irradiation requires the dose kernel libraries to produce dose maps. However, the Monte Carlo approach for dose kernel library generation is time- and space-consuming. To optimize spatiotemporal efficiency, we propose a machine learning approach based on 3D implicit neural functions to model focused kV X-ray dose kernels for small animal brain radiation. Based on the rat head phantom model, we define a library spanning a parameter space with proper range and resolution to cover the rat brain size and structure variation. We use the TOPAS-based Monte Carlo simulation for data preparation. Each dose kernel can be characterized by a depth parameter triplet. Implicit neural functions are trained to map spatial coordinates with a depth parameter triplet to their corresponding dose values. We find the periodic activation function of hidden neural layers and coordinate embedding of coordinates are important for accurate modeling. Experimental results show that we compress the training library from 17 GB to only 2.27 MB with high accuracy, and it is also shown to predict dose kernels in the testing dose kernels with an acceleration rate of 50,000 times, which is crucial to study the choice of optimal lens. This work is the first attempt to develop machine-learning techniques for modeling dose kernel libraries for small animal brain radiation.