Adarsh S Mcold start
000
resolving runtime01/05
← All work

Radiography

AI-assisted radiographic analysis

A radiographic analysis system reaching 95% accuracy, built to cut the time a clinician spends before reaching a diagnosis — not to replace the clinician.

Where
Fiverr — freelance
When
2024 — 2026
Role
Machine Learning Engineer
Stack
PyTorch · CNNs · Transfer learning · Medical imaging
95%Accuracyradiographic analysis
40%Diagnosis timereduction in manual review

Problem

Radiographic review is bottlenecked on expert attention, and that attention is spent evenly across a caseload where most studies are unremarkable. The opportunity is not autonomous diagnosis — it is getting the expert to the studies that need them faster.

Constraints

  • 01A false negative is not symmetric with a false positive in clinical use
  • 02The clinician stays in the loop; the system assists, it does not decide
  • 03Limited labelled data relative to general-purpose vision corpora
  • 04Output has to be interpretable enough for a professional to trust or overrule

Architecture

INTAKEPREPAREMODELREVIEWREPORTRadiographic studyimaging inputNormalisationcontrast, orientationAugmentationlimited label budgetCNN backbonetransfer learningConfidence scoringper-studyTriage orderingexpert attention firstClinician reviewalways in the loopReported findingclinician-signed
fig.01A convolutional pipeline with transfer learning doing the heavy lifting, and the clinician kept at the decision point rather than after it.

Key decisions & trade-offs

  • Assistive triage rather than autonomous diagnosis

    Framing the system as ordering the queue rather than answering the question is what made a 95% model useful. The clinician still reads every study they sign; they just read the ones that matter first.

    Trade-offIt caps the headline benefit — you cannot claim to have removed the review step. It is also the only framing that is deployable in a clinical setting.

  • Transfer learning against a small label budget

    Medical labels are expensive and scarce. Starting from a pre-trained vision backbone and adapting it means the limited labelled data is spent teaching the domain rather than teaching edges and textures from scratch.

    Trade-offThe model inherits whatever biases the source corpus carries, and natural-image pretraining is a genuinely different distribution from radiography. Augmentation narrows the gap; it does not close it.

Results

95%Accuracyheld-out radiographic studies
40%Manual diagnosis timereduction