Well the ways that don't involve surgery are themselves less accurate heuristics with false positives. How can they calibrate without being certain if it is a true or false positive or negative?

Better imaging is an option but that often falls into dangerous in itself or easier said than done like say trying to do detection of cancer by scent.

Machine learning is fundamentally GIGO and tends to learn smartass shortcuts if one isn't careful. "The healthy sample data is less zoomed in while the cancer ones are zoomed in - therefore full chest x-ray is fine but a lung X-ray is cancer."