Web-Based Deep Learning System for Evaluating Lu-PSMA Therapy Efficacy in Prostate Cancer Using PET/CT

This project presents an AI-powered theranostics platform designed to support physicians in evaluating treatment efficacy in metastatic prostate cancer (mPC) patients using Lu-PSMA PET/CT scans.

Developed by: Muhammed ElNajjar, Omar Shata, Abduallah Omran, Mohand Attya, Carole Bekhit Under Supervision of: Dr. Ahmed Ehab and Dr. Manar Nasser Faculty of Engineering, Cairo University – Department of Systems and Biomedical Engineering

---

Table of Contents

• Overview
• System Architecture
• Key Features
• AI Methodologies
• Results
• Limitations & Future Work
• Demo

---

Overview

Prostate cancer remains a leading cause of cancer mortality in men. Theranostics, combining diagnostics and targeted therapy, offers a modern solution. Our platform provides:

• AI-driven segmentation & classification
• Longitudinal tracking tools
• Web-based DICOM visualization
• Integration of ResNet & UNet models

---

System Architecture

The platform is composed of:

• PACS Server – Secure DICOM storage with compression & backups
• Study List – Interface to search, organize, and classify patient scans
• Web-based DICOM Viewer – Compare axial, sagittal, and coronal views + 3D MIP
• AI Module –
  • Prostate segmentation (Attention UNet)
  • Lesion classification (3D ResNet50)
  • Lesion segmentation (3D UNet w/ CT & PET input)

---

Key Features

• Interactive DICOM Viewer
  Zoom, pan, crosshair, 3D MIP, segmentation tools, slice comparison, and volume trend analysis.

• AI-Powered Tasks

• Prostate segmentation using Attention UNet

• Lesion classification with 3D ResNet

• Dual-modality lesion segmentation using 3D UNet

• Longitudinal Tracking
  Compare multiple time points, track treatment response via tumor volume and tracer uptake.

---

AI Methodologies

Prostate Segmentation

• Dataset: TCIA – 129 CT studies
• Model: 2D UNet, 3D UNet, Attention UNet
• Best Performance: Attention UNet (Dice = 0.85)

Lesion Classification

• Dataset: AutoPET Challenge
• Model: Custom 3D CNN vs ResNet50
• Best Performance: ResNet50 (Accuracy = 73%, Precision = 85%)

Lesion Segmentation

• Input: Dual-modality (PET + CT)
• Model: 3D UNet
• Dice Score: 0.67 (after post-processing)

---

Results

Task	Model	Score
Prostate Segmentation	Attention UNet	0.85 Dice
Lesion Segmentation	3D UNet	0.67 Dice
Lesion Classification	ResNet50	73% Accuracy

Post-processing improved lesion segmentation by 4% using body masking.

---

Limitations & Future Work

• Current models trained on FDG PET/CT datasets, not PSMA
• Plan to retrain on PSMA-labeled datasets from Misr Radiology Center
• Future improvements in segmentation accuracy with more advanced models

---

🎥 Demo

Watch the YouTube Demo

---