tbp-skin-detect

Skin Cancer Detection from 3D Total Body Photos

Project Duration: Aug 15, 2024 - Sep 7, 2024

---

🧠 Objective

TThe goal was to build binary classifiers to predict malignant skin lesions from single-lesion crops extracted from 3D total body photos (TBP). This project is part of the ISIC 2024 - Skin Cancer Detection from 3D-TBP Kaggle competition. Submissions were evaluated on partial AUC (pAUC) for true positive rates (TPR) above 80%.

---

🧩 Approach

You can explore the complete methodology in this notebook: 🔗 ISIC24 - Heavy Feature Eng with Polars + Boosting + CV + Ensemble Blend

Key steps followed:

• ✏️ Feature Engineering: spatial insights

  • Extracted 3D landmark distances using pairwise Euclidean computations across TBP points.
  • Constructed derived geometric features reflecting anatomical symmetry and patient-level spatial variation.

• ⚖️ Patient-Level Normalization: consistency modeling

  • Applied normalization of features at the patient level to control for inter-subject variability.
  • Included feature columns for image count per patient.

• 📊 Categorical Handling: info retention

  • Employed OneHotEncoder for categorical variables.
  • Converted them to category dtype for memory efficiency.

• 🧰 Ensemble Learning: reducing model variance

  • Trained LightGBM, XGBoost, and CatBoost models independently.
  • Combined predictions using a weighted ensemble method for improved pAUC.

• 📊 Custom Evaluation Metric: tailored pAUC

  • Implemented a custom scoring function to simulate competition-specific pAUC above 80% TPR.
  • This metric guided model selection and cross-validation.

---

🏆 Results / Outcomes

• ✅ Public Leaderboard Scores:

  • 0.18368, 0.18412, 0.18519

• 🏁 Private Leaderboard Scores:

  • 0.16733, 0.16753, 0.16930 (final best)

• 🥇 Rank Achieved:

  • Placed 184th out of 3410 participants and 2739 teams as a solo competitor

---

🔗 References

• 🏆 Kaggle Competition: ISIC 2024 - Skin Cancer Detection with 3D-TBP

---

🛠️ Tech Stack

• Language: Python 🐍
• Libraries:
  • polars for dataframe operations
  • pandas, numpy for numerical tasks
  • stratifiedKFold for cross validation
  • sklearn, lightgbm, xgboost, catboost for modeling
  • matplotlib, seaborn for visualization
  • optuna for hyperparameter tuning
• Tools:
  • Jupyter Notebook / Kaggle Notebooks 📓 for experimentation and code
  • Custom Python metric functions for pAUC evaluation