10. From what percent does “there is infiltration” start? — thresholds and names · Dr. Seok's Tumour-Infiltrating Lymphocytes, First Steps

Clinical pathology has an old grading used for melanoma: brisk / non-brisk / absent. "Lymphocytes have infiltrated vigorously / sparsely / not at all."

A pathologist assigns that grade by eye. We, on the other hand, have a continuous number. To carry it over we need a threshold — "5% or more is brisk", say.

Where did that 5% come from? From nowhere. Someone decided it, and someone else decides differently.

At a threshold of 0.5% even breast cancer is 73% "infiltrated"; at 20%, only 2.5% survives.

Python Runs in your browser

import til

d = til.slides()
cuts = [0.5, 1, 2, 5, 10, 20]

print("%-6s" % "type" + "".join("%8s%%" % c for c in cuts))
for t in til.TUMORS:
    g = d[d.tumor == t]
    share = [100 * (g.til_pct >= c).mean() for c in cuts]
    print("%-6s" % t + "".join("%8.1f" % v for v in share))
print()
print("Each number is '%% of slides called infiltrated at that threshold'.")

Press the threshold buttons. The ranking of cancer types mostly holds, but the absolute share swings wildly.

Loading…

Even as the threshold moves, the ranking of cancer types is fairly stable. Stomach and rectum stay near the top, breast and prostate near the bottom. That is good news — "which cancer type has more infiltration" is not very threshold-sensitive.

The bad news is the sentence about one patient. "This patient has infiltration" hangs entirely on one threshold — and, as lesson 11 will show, it can flip on a different slide from the same patient.

Of 405 melanoma patients, whether 190 or 297 are brisk is up to us.

Python Runs in your browser

import til

d = til.patients()          # one slide per patient
g = d[d.tumor == "SKCM"]    # melanoma - where the brisk grade is actually used

for cut in (2, 5, 10):
    brisk = g[g.til_pct >= cut]
    print("threshold %2d%% ->  brisk %3d of %d patients  (%4.1f%%)"
          % (cut, len(brisk), len(g), 100*len(brisk)/len(g)))
print()
print("Same patients. Only the threshold moved, and the count of 'brisk patients' moved a long way.")

🤔 Check yourself

What stays relatively stable as the threshold moves?

🤔 Check yourself

A paper says 'we defined TIL-high as 5% or above'. What should you ask?