Using Video2X (https://github.com/k4yt3x/video2x), I've done a little AI upscale quick test with 'waifu2x-ncnn-vulkan' cunet model (using only the upscale option, as denoising is useless here: the DOS version doesn't have mpeg1 compression artefacts, only the CDi version uses it).
Process:
DosBox internal lossless capture (320*200)
|-Video2X waifu2x-ncnn-vulkan AI upscale (640*400)
|-VirtualDub2 resize filter with nearest neighbor (1280*800)