Apppublic

build-small-hackathon/ct-app

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

2likes

App README

🩻 CT Report Generator

An automated 3D volumetric reporting pipeline for CT scans, powered by TotalSegmentator (3D U-Net) (⚡ ~30 Million Total Parameters) — deployed serverlessly on Modal.

📺 [Watch the full video demo and post on X (Twitter)!](https://x.com/AKIS23820044161/status/2066586748541657272)

Special Bonus Targets : Tiny Titan (~30M parameters model) · Off-Brand Award

📖 Overview

CT report generator is a Gradio-based clinical dashboard that automates the extraction and quantification of anatomical structures from 3D CT scans. It processes raw .nii / .nii.gz volumetric data, calculates the exact volume of dozens of internal organs, and automatically flags any measurements that fall outside of expected healthy reference ranges (e.g., hepatomegaly, splenomegaly, or asymmetrical kidneys).

🚀 Features

Feature	Description
🧠 Total Body Segmentation	Automatically identifies and segments major solid organs, thoracic structures, and GI/GU tracts.
📊 Clinical Volume Alerts	Cross-references organ volumes with normal adult reference ranges and flags anomalies (e.g. Enlarged liver, asymmetric lungs).
🖼️ Cross-Section Preview	Generates an immediate mid-axial visual slice of the uploaded 3D volume.
📑 PDF Report Generation	Automatically compiles the findings into a clean, professional, downloadable PDF clinical report using WeasyPrint.

🤖 AI Models Used

1. `TotalSegmentator` — 3D Anatomical Segmentation

—Architecture: 3D U-Net (nnU-Net framework)
—Total Parameters: ~30 Million (30M)
—Task: 3D medical image segmentation.
—Used for: Identifying and calculating the exact cubic centimeter (cm³) volume of 100+ anatomical structures from raw CT scans.
—Inference: Fast-mode enabled for rapid screening on Modal A10G GPU.

🏗️ Architecture

GRADIO FRONTEND (app.py)
  ├── 3D Visualization     → nibabel + PIL (Mid-axial slice rendering)
  ├── Validation           → Checks for valid 3D shape and intensity spread
  ├── PDF Generation       → WeasyPrint HTML-to-PDF conversion
  └── Remote RPC           → Connects to Modal backend via `modal.Cls`

MODAL SERVERLESS BACKEND (backend.py)
  └── Segmenter [A10G]     → `TotalSegmentator` subprocess
                           → JSON parsing & Reference Range Logic
                           → Returns structured clinical findings

🖥️ GPU Resources (Modal)

Container	GPU	Model(s)	Purpose
`Segmenter`	A10G (24GB)	TotalSegmentator 3D U-Net	Heavy volumetric segmentation and pixel quantification

⚙️ Setup & Deployment

bash

# 1. Install dependencies
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# 2. Deploy Modal backend
modal deploy backend.py

# 3. Run the Gradio frontend
python app.py