metadata

title: Phronesis Medical Report Generator
emoji: 🧠
colorFrom: green
colorTo: gray
sdk: gradio
app_file: app.py
pinned: false
short_description: 'REPORT GEN AND CLASSIFICATION MODEL '

🧠 Phronesis: Medical Image Diagnosis & Report Generator

Phronesis is a multimodal AI tool that classifies medical CT scan images (DICOM or standard formats) and generates diagnostic reports using a combination of video classification and medical language generation.

🚀 Demo

Upload a set of DICOM (.dcm, .ima) or image (.png, .jpg) files representing slices of a CT scan. The model will:

🏷️ Predict a class: acute, normal, chronic, or lacunar
📋 Generate a short radiology report

Live App →

🏗️ Model Architecture

Vision Backbone: 3D ResNet-18 pretrained on Kinetics-400
Language Head: BioBART v2 (pretrained biomedical seq2seq model)
Bridge Module: Custom ImageToTextProjector to align visual features with the language model
CombinedModel: Unified architecture for classification + report generation

🧪 Tasks

Image Classification: Categorizes brain CT scans into one of four classes.
Report Generation: Produces diagnostic text conditioned on image features.

🖼️ Input Format

Minimum 1, maximum ~30 image slices per scan.
Acceptable file formats:
- DICOM (.dcm, .ima)
- PNG, JPEG

The model will sample or pad the series to 16 frames for temporal context.

📦 Dependencies

This app uses:

torch
transformers
torchvision
huggingface_hub
pydicom
gradio
PIL, numpy

🔐 Notes

This demo loads a private model from the Hugging Face Hub. Set your HF_TOKEN as a secret for the space if needed.
Do not use for real clinical decisions – intended for research/demo only.

🙋‍♂️ Credits

Developed by @baliddeki

Model weights: baliddeki/phronesis-ml
Language model: GanjinZero/biobart-v2-base

📄 License

MIT or Apache 2.0 (add yours here)