MooseML/EfficientNet-Cancer-Detection

Histopathologic Cancer Detection - EfficientNetV2S & ViT-Hybrid

This repository contains models for detecting metastatic cancer in histopathologic images.

• EfficientNetV2S: A Baseline CNN-based model for local feature extraction.
• ViT-Hybrid: A Transformer-based model that learns global dependencies. Both models were trained on the Histopathologic Cancer Detection Kaggle dataset

Model Performance

• EfficientNetV2S
  • Accuracy: 93.59% (Private), 93.74% (Public)
  • AUC: 0.9774
• ViT-Hybrid
  • Accuracy: 95.07% (Private), 94.87% (Public)
  • AUC: 0.9791
• ViT-Hybrid + TTA (Test-Time Augmentation)
  • Accuracy: 96.50% (Private), 96.75% (Public)

Model Use

from huggingface_hub import hf_hub_download
from tensorflow.keras.models import load_model

Download EfficientNetV2S model

model_path = hf_hub_download(repo_id="MooseML/EfficientNet-Cancer-Detection", filename="efficientnet_cancer_model.h5")
model = load_model(model_path)

Download ViT-Hybrid model

model_path_vit = hf_hub_download(repo_id="MooseML/EfficientNet-Cancer-Detection", filename="ViT_hybrid_cancer_model.h5")
model_vit = load_model(model_path_vit)

Github and Kaggle Links for Full Training Pipeline

• Full Training Code: GitHub Repository
• Kaggle Competition: Histopathologic Cancer Detection

---

license: mit