AI-Based Liver Diagnosis System from CT & MRI
Budget: ₹37,500 – ₹75,000 INR
1. Project Overview
Develop a clinically accurate AI system for automatic liver segmentation, lesion detection, lesion classification, and liver disease diagnosis using CT and MRI scans. The system must produce structured clinical reports and visual heatmaps, with an AUC ≥ 90%, and integrate into a React-based UI using Google Cloud Storage (GCS).
2. Scope of Work
Modality Support:
• Input: DICOM files
• Supported scans:
• CT (single-phase, multiphase: arterial, venous, delayed)
• MRI (T1, T2, DWI, DCE sequences)
Functional Modules
2.1. Preprocessing
• Parse DICOMs, detect modality & sequence type
• Resample, normalize, and prepare data
• Phase tagging for CT; Sequence tagging for MRI
2.2. Liver Segmentation
• Train 3D U-Net/nnU-Net to segment the liver
• Output: Binary mask (NIfTI)
• Target Dice score: ≥ 0.90
2.3. Lesion Detection
• Identify liver lesions: HCC, cysts, metastases, etc.
• Output: bounding box/mask per lesion
• Lesion sensitivity ≥ 85% (for lesions ≥ 10mm)
2.4. Lesion Classification
• Label each lesion: HCC, cyst, metastasis, hemangioma, FNH
• Use multiphase CT or multi-sequence MRI
• AUC ≥ 0.90 per class
2.5. Liver Disease Classifier
• Detect and classify:
• Fatty liver (mild/moderate/severe)
• Fibrosis stage (F0–F4)
• Cirrhosis (structural indicators)
2.6. Clinical Rule Engine
• Implement logic for:
• LI-RADS scoring (CT/MRI)
• Fibrosis staging
• BCLC staging (for HCC)
2.7. Explainability (Heatmaps)
• Generate Grad-CAM/SHAP overlays for:
• Lesion classification
• Disease classifier
• Output aligned with DICOM slices
2.8. Frontend Viewer (React)
• Upload DICOM to GCS
• Display liver masks, lesions, heatmaps
• Support doctor override/edit
• Trigger structured report downloads
2.9. Report Generator
• Output:
• JSON (structured report)
• PDF (printable clinical summary)
• Optional: DICOM-SR
• Include all findings, scores, and overlays
3. Privacy, Security & Compliance
PHI Protection:
• DICOM tag scrubber to anonymize patient data
GCS Integration:
• Use IAM roles or token-based access to secure uploads and retrieval
Open-source Stack Only:
• Use only open-source frameworks (e.g., PyTorch, MONAI, OHIF)
• No proprietary SaaS, paid APIs, or closed-source dependencies allowed
NDA Required:
• Vendor must sign a Non-Disclosure Agreement (NDA) before project initiation or data access
4. Deliverables
Item Format
Trained models .pth / ONNX format
Liver & lesion masks NIfTI
Final labeled dataset DICOM + COCO JSON or CVAT XML + CSV metadata
Inference script or Docker container CLI or optional REST API
Heatmap overlays PNG + coordinates in JSON
React UI Source + Dockerfile
Test suite PyTest or equivalent
Reports JSON + PDF (and optional DICOM-SR)
Documentation Markdown + PDF (user + developer guide)
5. QA & Testing
• Vendor must submit 25 sample cases with model outputs and heatmaps for radiologist spot check
• Client reserves the right to introduce test scans or bring in a radiologist for manual review
6. Performance Requirements
Metric Target
Liver segmentation Dice ≥ 0.90
Lesion detection ≥ 85% sensitivity (10mm+)
Lesion classification AUC ≥ 0.90 per class
Disease diagnosis Stage prediction accuracy ≥ 85%
Report accuracy JSON/PDF export must be parseable and match results
7. Model Deployment
• Vendor must provide:
• Inference-ready Docker container
• CLI script or FastAPI/Flask-based REST endpoint
• README with input/output example usage
8. IP Ownership
All source code, trained models, labeled data, frontend UI, and documentation developed under this contract are the sole intellectual property of the client and must be handed over at the end of the project. Vendor relinquishes all commercial claims.
9. Retraining & Failure Handling
• If AUC < 90% or lesion detection performance is subpar, vendor agrees to:
• Perform up to 2 rounds of retraining or optimization at no additional cost
10. Timeline (14 Weeks)
Phase Weeks
Preprocessing, GCS setup Week 1–3
Liver segmentation model Week 2–5
Lesion detection/classification Week 4–8
Disease classification + rules Week 6–10
Heatmaps + React UI Week 9–12
Testing, reporting, delivery Week 12–14
Develop a clinically accurate AI system for automatic liver segmentation, lesion detection, lesion classification, and liver disease diagnosis using CT and MRI scans. The system must produce structured clinical reports and visual heatmaps, with an AUC ≥ 90%, and integrate into a React-based UI using Google Cloud Storage (GCS).
2. Scope of Work
Modality Support:
• Input: DICOM files
• Supported scans:
• CT (single-phase, multiphase: arterial, venous, delayed)
• MRI (T1, T2, DWI, DCE sequences)
Functional Modules
2.1. Preprocessing
• Parse DICOMs, detect modality & sequence type
• Resample, normalize, and prepare data
• Phase tagging for CT; Sequence tagging for MRI
2.2. Liver Segmentation
• Train 3D U-Net/nnU-Net to segment the liver
• Output: Binary mask (NIfTI)
• Target Dice score: ≥ 0.90
2.3. Lesion Detection
• Identify liver lesions: HCC, cysts, metastases, etc.
• Output: bounding box/mask per lesion
• Lesion sensitivity ≥ 85% (for lesions ≥ 10mm)
2.4. Lesion Classification
• Label each lesion: HCC, cyst, metastasis, hemangioma, FNH
• Use multiphase CT or multi-sequence MRI
• AUC ≥ 0.90 per class
2.5. Liver Disease Classifier
• Detect and classify:
• Fatty liver (mild/moderate/severe)
• Fibrosis stage (F0–F4)
• Cirrhosis (structural indicators)
2.6. Clinical Rule Engine
• Implement logic for:
• LI-RADS scoring (CT/MRI)
• Fibrosis staging
• BCLC staging (for HCC)
2.7. Explainability (Heatmaps)
• Generate Grad-CAM/SHAP overlays for:
• Lesion classification
• Disease classifier
• Output aligned with DICOM slices
2.8. Frontend Viewer (React)
• Upload DICOM to GCS
• Display liver masks, lesions, heatmaps
• Support doctor override/edit
• Trigger structured report downloads
2.9. Report Generator
• Output:
• JSON (structured report)
• PDF (printable clinical summary)
• Optional: DICOM-SR
• Include all findings, scores, and overlays
3. Privacy, Security & Compliance
PHI Protection:
• DICOM tag scrubber to anonymize patient data
GCS Integration:
• Use IAM roles or token-based access to secure uploads and retrieval
Open-source Stack Only:
• Use only open-source frameworks (e.g., PyTorch, MONAI, OHIF)
• No proprietary SaaS, paid APIs, or closed-source dependencies allowed
NDA Required:
• Vendor must sign a Non-Disclosure Agreement (NDA) before project initiation or data access
4. Deliverables
Item Format
Trained models .pth / ONNX format
Liver & lesion masks NIfTI
Final labeled dataset DICOM + COCO JSON or CVAT XML + CSV metadata
Inference script or Docker container CLI or optional REST API
Heatmap overlays PNG + coordinates in JSON
React UI Source + Dockerfile
Test suite PyTest or equivalent
Reports JSON + PDF (and optional DICOM-SR)
Documentation Markdown + PDF (user + developer guide)
5. QA & Testing
• Vendor must submit 25 sample cases with model outputs and heatmaps for radiologist spot check
• Client reserves the right to introduce test scans or bring in a radiologist for manual review
6. Performance Requirements
Metric Target
Liver segmentation Dice ≥ 0.90
Lesion detection ≥ 85% sensitivity (10mm+)
Lesion classification AUC ≥ 0.90 per class
Disease diagnosis Stage prediction accuracy ≥ 85%
Report accuracy JSON/PDF export must be parseable and match results
7. Model Deployment
• Vendor must provide:
• Inference-ready Docker container
• CLI script or FastAPI/Flask-based REST endpoint
• README with input/output example usage
8. IP Ownership
All source code, trained models, labeled data, frontend UI, and documentation developed under this contract are the sole intellectual property of the client and must be handed over at the end of the project. Vendor relinquishes all commercial claims.
9. Retraining & Failure Handling
• If AUC < 90% or lesion detection performance is subpar, vendor agrees to:
• Perform up to 2 rounds of retraining or optimization at no additional cost
10. Timeline (14 Weeks)
Phase Weeks
Preprocessing, GCS setup Week 1–3
Liver segmentation model Week 2–5
Lesion detection/classification Week 4–8
Disease classification + rules Week 6–10
Heatmaps + React UI Week 9–12
Testing, reporting, delivery Week 12–14