Pathologist here (doctor who diagnoses cancer), who happens to have a degree in physics and works on diagnostic ML problems, including cancer.
Cancer is not a simple anomaly detection problem. I mean, there's that, but so much more. These cells, each one of them, by their very nature, looks exactly like your normal healthy cells, on the outside. It's what's inside the cell that's going to kill you. The uncontrolled replication. There are 36 trillion cells in the human body. How are you going to monitor them all? Well, turns out we have several methods built in, collectively called the immune system. But again, they're mainly looking for "not self". Because if they were looking for "self" you'd have another problem, called auto-immunity.
The uncertainty is real. It's not a hypothetical uncertainty. Combinatorics is a bitch.
There are 3 billion base pairs per normal human cell. The difference between a normal cell and a cancer cell can be on a similar order (a cancer cell may have many billions or less than a billion base pairs). There are similar problems for the number of proteins, lipids, polysaccharides, metal ions etc, per cell.
Three billion times 36 trillion, oh, and many generations of many of those 36 trillion over time. So, let's casually say a billion billion cells in a human life time. Times 3 billion base pairs. If one of those cells gets out of control, you've got a problem. Shockingly, only one in six people die of cancer.
Only under the most austere circumstances can we partially characterize a single cancer cell, and even then we waste many, many other cancer cells to surface that one cell (e.g. single cell transcriptomics).
Other methods allow us to examine many cells, but we can't examine them as closely (histology, histochemistry, immunochemistry, in situ hybridization, flow cytometry, targeted genomics, shotgun sequencing, karyotyping, etc), and we still never see most of the cells.
To get a basic understanding, it's advisable to take the ground up approach used in statistical mechanics: in cancer, from the ground up, a single cell is the source of the initial problem. That cell and its progeny divide many times, let's say 30-40 times. Now you have a billion cells, maybe 10-100 billion cells. Every one of which is starting from an unmanaged state, highly vulnerable to additional mutations. And probably the cell of origin died 20 generations before you find the tumor. Even in a basic science research setting, it would be exceedingly challenging to demonstrate you had found "the cell of origin".
This is very similar to physics: there are things we can know at one energy level that we can't know at another energy level. You can't explore Bose-Einstein condensates with the LHC. You're off by 20 orders of magnitude. You can't do single cell transcriptomics on a 1 kg en bloc cancer resection specimen, you're off by 20 orders of magnitude.
Complicating matters, fission is actually pretty straight forward compared to biology. In fission, you've got a very small number of elements involved, at very high, specified purity. In biology, you can barely guess most of what the organism consumed in the last 24 hours, let alone what they've been exposed to over a lifetime.
Cancer, nuclear physics, internet-scale computation, most of the really interesting problems: you can't just "take pictures" of the whole thing. It would be like assuming you can "just understand" what's happening in an actual nuclear explosion using some cameras and a sound understanding of math. Or the proposal that we could just understand the global economy by examining the ledger of all transactions. It's ridiculously beyond the realm of possible.
Even in the Trinity explosion, a highly controlled, intensively studied nuclear explosion, we can't even be sure how many neutrons the beryllium-polonium initiator produced. 8? 10? Not really sure which atoms produced them, for sure. And how many got produced in each succeeding fission generation? Meh? I mean, we can do some statistics, but that's it. That's roughly the scale of the problem we're dealing with in cancer research: where'd the thing come from? And where's it going? Statistically, we can make some guesses, but no one understands the whole thing.