Lung SBRT Prediction Model
Budget: $250 – $750 USD
I have a curated Lung SBRT dataset containing patient CT images, target/OAR contours, and corresponding clinically approved dose distributions, supplemented by patient demographic and treatment parameters. All data are stored as organized database records linking to DICOM files. My goal is to build a proof-of-concept Convolutional Neural Network (CNN) that can predict 3D dose distributions and automatically suggest planning objectives for new SBRT cases.
Scope
• Data pipeline – extract, preprocess, and split the DICOM-linked database while preserving PHI safeguards. This includes CT normalization, resampling, and structure mask generation.
• Model development – design and train a 3D CNN (U-Net–based) architecture tailored to volumetric medical imaging inputs; include hyperparameter tuning, class balance adjustments, and cross-validation.
• Evaluation – report MAE, DVH metrics, and 3D γ-analysis, with visual heatmap comparisons between predicted and clinical dose distributions to assess clinical relevance.
• Deliverables – commented Python code (TensorFlow or PyTorch), reproducible Jupyter notebook, inference pipeline, and a concise technical report summarizing model architecture, performance, and potential future improvements.
I can provide secure access to the anonymized database and a sample of existing planning parameters to facilitate feature extraction and model initialization. Please confirm your preferred environment or GPU setup so we can configure the workflow efficiently.
Scope
• Data pipeline – extract, preprocess, and split the DICOM-linked database while preserving PHI safeguards. This includes CT normalization, resampling, and structure mask generation.
• Model development – design and train a 3D CNN (U-Net–based) architecture tailored to volumetric medical imaging inputs; include hyperparameter tuning, class balance adjustments, and cross-validation.
• Evaluation – report MAE, DVH metrics, and 3D γ-analysis, with visual heatmap comparisons between predicted and clinical dose distributions to assess clinical relevance.
• Deliverables – commented Python code (TensorFlow or PyTorch), reproducible Jupyter notebook, inference pipeline, and a concise technical report summarizing model architecture, performance, and potential future improvements.
I can provide secure access to the anonymized database and a sample of existing planning parameters to facilitate feature extraction and model initialization. Please confirm your preferred environment or GPU setup so we can configure the workflow efficiently.