Disease Classification from Scans
Disease classification sorts whole scans into diagnostic buckets where prevalence and calibration matter more than raw accuracy.
Why Does This Exist?
Triage needs sort, not outline: normal versus urgent across thousands of X-rays. A percent accurate model fails when disease hits percent and every positive needs review. Input is a full scan; output is calibrated probabilities plus a referral cutoff. Localization helps trust but the decision is scan-level.
Think of It Like This
Sorting mail into urgent versus bulk bins
Postal workers sort thousands of letters by stamp and weight, not by reading each one. Urgent bins stay small because most mail is bulk. Misreading bulk as urgent jams the express truck.
Scans are letters and cutoffs are bin labels. The analogy stops at calibration: bins need true probabilities, not just order.
How It Actually Works
A backbone pools to per disease with binary cross-entropy and label smoothing. Multi-label heads tag chest findings independently. Calibration plots predicted against true rates; temperature scaling fixes overconfidence. Cutoffs set sensitivity near for rule-out versus specificity for referral.
Worked example
One thousand chest films hold pneumothoraces. The net at cutoff catches with false alarms, so review load is and miss count . Accuracy is but utility is misses. Raising to cuts alarms to but misses rise to . Prevalence picks the point, not accuracy.
Watch Out For
Site prevalence shift breaking cutoffs
A cutoff tuned at percent prevalence alarms twice as often at percent in winter. The symptom is sudden review floods. Recalibrate cutoffs per site seasonally, not once globally.
The Quick Version
- Classification sorts scans for triage, not for outlining disease.
- Low prevalence makes calibration and cutoffs dominate accuracy.
- Multi-label heads tag each finding independently.
- Do not ship one global cutoff; tune per site and season.
- Report misses at the chosen point, not accuracy alone.