CephaloTrain AI Studio — Cephalometric Dataset, Local Qwen/Vision Trainer & Evaluation Lab

Tip: Drag any landmark point directly on the X-ray canvas to refine its ground-truth coordinates, then click "Save Ground Truth". X: -- , Y: --

Steiner & Tweed Diagnosis

Class I

SNA Angle (Maxilla)

--°

Norm: 82.0° ± 2°

SNB Angle (Mandible)

--°

Norm: 80.0° ± 2°

ANB Angle (Sagittal)

--°

Norm: 2.0° ± 2°

FMA (Tweed Vertical)

--°

Norm: 25.0° ± 3°

IMPA (Lower Incisor)

--°

Norm: 90.0° ± 5°

FMIA Angle

--°

Norm: 65.0° ± 5°

23 Anatomical Landmarks (px)

23 / 23 Placed

Model Training Configuration

CPU Optimized

Model Checkpoint Name

Local AI Architecture (Optimized for Xeon 4-Core / 16GB RAM)

Training Epochs

Learning Rate

MLP Hidden Units

On-the-Fly Aug Factor

Template Patch Radius (px)

ROI Search Window (px)

Sync with Qwen / Ollama (`cephalotrace-qwen`)

Automatically compiles the annotated dataset into an Ollama Modelfile + Qwen2-VL JSONL instruction dataset during training.

How Local Training Works on Your Machine:
1. Extracts CLAHE & Sobel multi-scale gradient features from all training X-rays.
2. Trains the Global Neural Coordinate Regressor + PCA Active Shape Model.
3. Builds a high-resolution 23-landmark Normalized Cross-Correlation (NCC) template bank.
4. Exports Qwen2-VL JSONL & custom Ollama Modelfile.

Training Engine Ready

Select hyperparameters and click "Start Local AI Training".

Epoch 0 / 0

-- mm

Val Mean Radial Error (MRE)

-- px

Val Radial Error (Pixels)

--%

SDR @ 2.0 mm (Clinical Pass)

--%

SDR @ 4.0 mm

Training & Validation Loss (MSE) Cyan: Train | Amber: Val

Validation MRE (mm) & SDR@2mm (%) Green: SDR@2mm | Rose: MRE

Per-Landmark Localization Precision (Mean Radial Error in mm across 23 Landmarks) Dashed Line = 2.0 mm Clinical Threshold

Model Test & Benchmark

1. Select Trained AI Model

2. Select Dataset X-Ray to Test

Or Upload Unseen Test X-Ray / DICOM 📂 Upload External X-Ray for Testing

Visual Evaluation Overlays

● AI Predicted Landmarks (Cyan) ● Ground Truth Landmarks (Green) ↗ Error Displacement Vectors ◎ Confidence Heatmaps 📐 Steiner & Tweed Tracing Lines

Batch Cohort Benchmark

Evaluate the selected model across all Validation & Test X-rays to compute aggregate MRE, SDR, and Skeletal Classification accuracy.

🧠 Local Qwen / Ollama Clinical Cephalometric Report CephaloTrace-Qwen Engine

            Select an X-ray and click "Run AI Prediction & Evaluate" to generate landmark coordinates, error metrics, and clinical report.
          

Ground Truth vs. AI Prediction

Cephalometric Angles Comparison

Measurement	Ground Truth	AI Predicted	Δ Diff

23 Landmarks Radial Error Breakdown

ID	Pred (x,y)	Err (px)	Err (mm)	Status

Local Ollama & Qwen Runtime Status

Checking local Ollama service...

Build Custom cephalotrace-qwen Ollama Model from Dataset

Recommended Local Models for Xeon 8573C (16GB RAM)

Drop-In REST API for CephaloTrace

Your trained model is automatically served on POST http://127.0.0.1:5050/predict, 100% compatible with your Trace and Trace-AI applications:

curl -X POST http://127.0.0.1:5050/predict \ -F "image=@dataset/images/sample1.jpg"

Generated Qwen2-VL & Ollama Training Artifacts

Inspect or download the compiled fine-tuning files generated from your annotated cephalometric X-ray dataset.

exports/Modelfile ⬇ Download File

Loading artifact preview...