← Back to Exhibition
AI & Machine LearningApr 2026 — Jun 2026

Colon Histopathology AI Classifier

Machine Learning Engineer & Data Scientist
Scroll
01 — Overview

End-to-end ML research pipeline automating colon cell cancer detection and cell-type classification from 27×27 RGB histopathology patches. Rigorous patient-level evaluation on unseen clinical data.

Exceeded diagnostic targets (≥0.90 cancer detection, ≥0.60 cell-type) on genuinely unseen patients — not just new images from seen patients.

02 — Gallery
03 — Key Contributions
01

Designed, trained, and evaluated 12 distinct models across two diagnostic tasks using classical ML (SVM, Random Forest) and custom PyTorch CNNs.

02

Implemented Test-Time Augmentation (TTA), dynamic learning rate scheduling, and per-epoch data augmentation.

03

Conducted controlled transfer learning experiments with CIFAR-10 pretraining, documenting negative transfer between natural-image and medical domains.

04

Architected strict patient-level train/validation/test splits preventing data leakage from staining and scanner artifacts.

05

Achieved 0.91 Macro F1 on binary cancer detection and 0.75 Macro F1 on 4-class cell-type classification on held-out unseen patients.

06

Performed comprehensive EDA using PCA, t-SNE, and K-Means clustering to analyze latent space and validate decision boundaries.

04 — Features
01

Patient-level evaluation preventing clinical data leakage

02

12 models across cancer detection and cell-type classification

03

CNN + transfer learning + TTA optimization

04

Macro F1 as primary metric for class imbalance

05

Controlled experiments isolating augmentation and ensemble effects

05 — Links