Lesion Detection in Medical Scans
Lesion detection flags small abnormal spots with boxes or points so radiologists check them before they grow.
Why Does This Exist?
Cancers start as mm spots in slices. Humans tire after dozens of scans. Object detection ports to 3D, but medical errors cost biopsies or missed cures. Input is a CT or X-ray volume; output is boxes plus scores tuned for screening recall. Two false alarms per scan exhaust readers; ten make the tool mute.
Think of It Like This
Airport screeners flagging small blades in bags
Screeners scan hundreds of bags for centimeter blades. Missing one is disaster, but flagging ten clean bags per hour stalls the line. Supervisors tune sensitivity to catch blades while keeping flags near two per hour.
Lesions are blades and scans are bags. The analogy stops at harm: bag flags delay while scan flags biopsy.
How It Actually Works
A 3D backbone proposes spots at mm stride with focal loss on -to- foreground odds. Non-maximum suppression merges 3D overlaps at IoU . Operating points fix sensitivity near for mm nodules; specificity follows from false positives per scan.
Worked example
A chest CT holds one mm nodule plus vessel crossings. The net scores the nodule and the top vessel . Threshold catches the nodule with false alarm in this scan and mean across the test set. Dropping to lifts sensitivity to but false alarms triple to , past the usable line.
Watch Out For
Vessel crossings mimicking nodules on one plane
Axial cuts turn crossing vessels into round dots identical to nodules. The symptom is confident false boxes on every scan. Confirm roundness on coronal plus sagittal before flagging; vessels tube out while nodules stay round.
The Quick Version
- Lesion detection trades sensitivity against false alarms per scan.
- Five-millimeter spots need 3D context, not single slices.
- Two false alarms per scan is the practical ceiling for readers.
- Do not deploy at maximum sensitivity; pick the screening point.
- Vessel crossings are the top false friend on axial cuts.