The process of calculating 3D images or rendering as its called doesn’t always produce correct results. The sacrifice for faster rendering speed is more noise or a less accurate simulation. The idea is simple. During the process of rendering. When the computer is calculating the pixels. Why not incorporate user input. With my eye I can see where the problem areas are and suggest perhaps more samples in that area. For animations you would probably need some sort of machine learning but for still images I think the idea could be worth implementing.
Below I marked the problem areas to be rendered with 100 times more samples during the render time. It could also perhaps work as a render mask. So you can can store the problem areas for later use.