license: cc-by-nc-4.0 base_model: google/resnet-50 tags: - digital-pathology - camelyon16 - cnn - mil - baseline
BTRUST Co.'s CNN-MIL Baseline (CAMELYON16)
📌 Project Overview
This repository contains the CNN-based Multiple Instance Learning (MIL) model used as the performance baseline for our pathology AI research.
🏆 Institutional Achievement
Developed as part of National HPC Supporting Program by AICA, Gwangju, s.Korea and also partly through
National NPU Support Program by NIPA, Jincheon, s. Korea. Elice Group Co., Ltd. in Seoul, s. Korea
kindly provided an A100 x 2 GPUs based service for the model training.
This model represents our commitment to reducing the manual workload of pathologists through high-performance AI.
📊 Model Details
- Architecture: ResNet50-Backbone with Attention-based MIL Aggregator
- Training Data: CAMELYON16 (H&E Stained Slides)
- Framework: Keras / TensorFlow
- Target: Lymph node metastasized breast cancers
- Note: keras_hub utilizes standardized Vision Transformer weights originally researched and released by the Google/timm teams. The base_model tag on Hugging Face is used for lineage tracking. By pointing to the timm repository, your model will correctly show up as a "Derivative Model," which increases your visibility in the global AI ecosystem.
📁 Dataset & Data Availability
The model was trained on a curated version of the CAMELYON16 dataset, processed into multi-scale patches and masks.
Dataset Components:
- Tissue Masks: Automated tissue detection at 2.5x/10x.
- Tumor Masks: Expert-verified ground truth masks.
- Patches: Extracted at 2.5x (contextual) and 10.0x (morphological) magnifications.
Access:
Due to the significant storage size and ongoing curation for commercial spin-off readiness, the processed dataset is not publicly hosted at this time.
- Academic Researchers: Available upon reasonable request for validation purposes.
- Inquiries: Please contact [dskim@btrust.co.kr] for data access requests.
📊 Dataset Pipeline
We provide the full pipeline to convert original CAMELYON16 TIFF images into the TFRecord format used for training this model. Available at https://github.com/kimdesok/ViT-backbone-MIL-on-CAMELYON16/Convert_TIFs_to_TFRecords.ipynb
Data Components
- Source: Original CAMELYON16 WSIs (.tif)
- Output: Multi-scale TFRecord sets (2.5x and 10.0x magnification)
- Contents: Tissue masks, Tumor masks, and Patch sets.
Accessing the Data
The processed TFRecord files are hosted on our secure institutional storage due to their large scale.
- Scripts: See here for the TIFF-to-TFRecord conversion code.
- Download: To request access to the pre-processed TFRecord sets, please fill out our Data Request Form/Email us.
📊 Comparison Strategy
This model serves as the control to evaluate the performance gains achieved by our ViT_MIL and a newer ViT-MIL (Virchow-based) architecture.
| Metric | CNN-MIL (This Model) | ViT-MIL (Next-Gen) |
|---|---|---|
| Backbone | ResNet-50 (CNN) | ViT-Tiny/Virchow 2 |
| AUC | 0.9424 (Baseline) | no data |
| Inference Speed | 657 images/sec(A100x2) | no data |
🔗 Related Models
- Vit-MIL Model: kimdesok/vit_mil_camelyon16
🛠 Model Details
- Backbone: ResNet-50 (Pre-trained on ImageNet)
- Aggregator: Attention-based MIL
- Dataset: CAMELYON16 (10.0x TFRecords)
⚠️ License & Commercial Use
This model is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).
- Academics: Free to use for research and publications.
- Industry/Commercial: Use for-profit requires a separate commercial license.
- Inquiries: Please contact [dskim@btrust.co.kr] for licensing and collaboration.