Ariana DakaResumeContact
All work
ShippedCourse

RL for Skin Cancer Decision Support

Replication and extension of a Nature Medicine reinforcement-learning model that treats missed melanoma as costlier than over-treatment.

Role
Experimental replication
Team
3 people
When
Jan 2026 – May 2026
Context
NJIT Reinforcement Learning course
Stack
  • Python
  • TensorFlow
  • DQN
  • HAM10000
Code (opens in a new tab)Write-up (opens in a new tab)
ASYMMETRIC COST · ILLUSTRATIVEPREDICTEDACTUALTPFNFPTNCOSTLIESTEPISODESREWARD

Highlights

  1. 01Reproduced the paper's melanoma recall and confusion matrices on HAM10000 (10,015 images, 7 classes)
  2. 02Replicated the lesion-level management pipeline (dismiss / cryo / excise) and implemented the naive and threshold-SL baselines myself, enabling a true three-way comparison
  3. 03Wrote the evaluation plotting (confusion heatmaps, reward-convergence curves, paper-vs-ours overlays) and led the presentation storyline

Write-up

Problem

Classifiers optimized for accuracy treat every error the same, but in skin-cancer triage a missed melanoma costs far more than an unnecessary excision.

What I built

With a team of three, I reproduced the paper's melanoma recall and confusion matrices on 10,015 images, implemented the naive and threshold baselines, built the evaluation plots and led the presentation.

Next project