The other day, while examining a routine hematoxylin and eosin (H&E) slide—the most common stain used in pathological diagnosis—I felt a subtle sense of discomfort. Something did not quite fit.
That small “unease” eventually led to a diagnosis that would completely change the patient’s treatment plan.
Looking at the H&E section, I had the impression that the case did not follow the usual pattern. Rather than staying on the expected diagnostic track, I began to suspect a different disease altogether. To explore that possibility, I performed immunohistochemistry—a method used to identify specific proteins within tissue.
The result confirmed that my sense of discomfort had been correct. The diagnosis was established. Fortunately, it was a condition in which the treatment strategy would change 180 degrees depending on whether it was present or not.
But what struck me most happened afterward.
Once the diagnosis was confirmed by immunohistochemistry, I returned to the original H&E slide. This time, the disease was unmistakable. The abnormal cells expressing that protein seemed to leap into view.
“Were you all here the whole time?” I almost wanted to say.
It reminded me of those online puzzles in which you move a single matchstick to make an incorrect arithmetic equation correct. Before you know the answer, it feels nearly impossible. Once you see it, however, it becomes obvious.
Why does everything suddenly become clear once you know the answer?
Of course, if you do not know the disease at all, there is no starting point. That would be like attempting the arithmetic puzzle without knowing how addition or subtraction works. But beyond that basic knowledge, something interesting happens in the diagnostic process.
As I once discussed with AI, pathological diagnosis begins in a state of exploration. We look at tissue without knowing what disease is present and search for patterns. When a diagnosis is established, it means the disease has been recognized. And once recognized, a pattern forms.
That is why common diseases—such as gastric cancer, colorectal cancer, or breast cancer—can be diagnosed almost instantly. We have encountered them countless times.
Rare diseases are different. By definition, they are rare. Over ten years, one might diagnose 10,000 cases of common cancers, while encountering only two or three cases like the one I saw the other day. It is difficult to build stable patterns from such infrequent exposure.
Still, the diagnostic process itself is fascinating. Under the microscope, all the information initially appears fragmented and unstructured. Yet once various techniques are applied and a diagnosis is secured, the scattered pieces suddenly organize themselves into a coherent structure. A pathway to the diagnosis becomes visible.
In theory, once that pathway is learned, the next case should be easier. In reality, if you encounter such a case only once every two or three years, memory fades.
Excellent pathologists, however, seem to retain a vast number of such patterns. They have many “drawers” to pull from. My own drawers are not particularly numerous, although having seen many rare diseases, I know this field reasonably well. Even after more than thirty years in pathology, I still feel that my training is incomplete. The depth of medicine is profound.
People often say that AI will eventually replace pathologists. It is true that AI excels at pattern recognition. If trained on an enormous number of images, it can achieve high accuracy in diagnosing common diseases.
However, pathology is not merely image classification.
It requires the ability to sense subtle “unease,” to question whether a case is diverging from the expected path, and to reconstruct the diagnosis into a meaningful narrative once it is established. That is where the human role still lies.
AI reorganizes existing knowledge. It does not yet originate hypotheses from a vague sense of discomfort.
That said, with sufficient training data, AI can recognize patterns of common diseases, and such systems are already being implemented. A portion of the pathologist’s workload may indeed be delegated.
In a time of pathologist shortages, that may be, in some ways, welcome news.
