AI-powered digital telepathology workflow using OpenFlexure Microscopy, cloud integration, and ResNet50V2 for affordable cancer diagnosis in rural Kenya.
# Low-Cost Digital Telepathology Workflow for Cancer Diagnosis Using AI
This project presents a practical, low-cost digital telepathology system developed and deployed in rural Kenya to support cancer diagnosis through AI-assisted histology slide analysis. The system integrates 3D-printed OpenFlexure Microscopes, cloud-based image sharing (MyPathomation™), and a fine-tuned deep learning model (ResNet50V2) for image classification.
## 🎯 Objectives
- Develop a low-cost digital workflow for cancer diagnosis using open-source tools.
- Capture and digitize histology slides using OpenFlexure Microscopes.
- Enable remote diagnosis through cloud-based slide sharing (MyPathomation™).
- Train and evaluate AI models to classify histological images as benign or malignant.
- Validate AI predictions against expert pathologists and assess real-world feasibility.
## 🧠 Model
- **Architecture:** ResNet50V2 (transfer learning)
- **Dataset:** BreakHis (5,429 malignant and 2,480 benign breast cancer images)
- **Training:** 8 epochs with fine-tuning
- **Evaluation Metrics:**
- Accuracy: 94%
- Sensitivity: 90%
- Specificity: 97%
- ROC-AUC: 0.99
- Pathologist agreement: 60%
## 🛠 Workflow Overview
1. **Microscope Fabrication**
- 3D-printed OpenFlexure Microscopes using PLA and Raspberry Pi components.
2. **Slide Digitization**
- Histological slides scanned and stitched into high-resolution images.
3. **Cloud Upload & Review**
- Digitized slides uploaded to MyPathomation™ for remote review by trained pathologists.
4. **AI Model Training & Testing**
- ResNet50V2 trained on BreakHis dataset and evaluated using ROC-AUC, sensitivity, and specificity.
5. **Clinical Deployment**
- Deployed in a rural clinic in Kenya with real-time pathologist training and use.
## 📍 Key Features
- Low-cost, open-source telepathology platform
- Combines hardware, cloud services, and AI into a deployable pipeline
- High performance metrics despite hardware and bandwidth constraints
- Adaptable to other d …