1. Identifying the Problem: Slow & Inconsistent Diagnosis

Approach

To validate feasibility and user needs, we conducted data analysis, ML prototyping, and UX research to design an effective AI-driven solution.

Problem Overview

• COVID-19 spreads rapidly and can cause severe respiratory complications. The current gold standard for diagnosis, RT-PCR testing, has major limitations:

  • Slow results: Takes 24+ hours to return findings.

  • Inaccuracy: 30-70% sensitivity, leading to false negatives and delayed treatment.

  • Limited access: Requires laboratories and trained personnel, making it difficult to scale testing in overwhelmed healthcare systems.

User Pain Points

• Delays in treatment & isolation

• False negatives increase spread

• Limited access to rapid testing

Competitive Gaps

• RT-PCR is slow and resource-heavy.

• Existing imaging tools lack AI automation for detection.

Opportunity

• AI-powered imaging can provide faster, scalable, and automated COVID-19 detection, improving accuracy and accessibility.

2. Developing the AI Model: Feasibility & Implementation

Technology Feasibility

• AI in medical imaging: Research shows machine learning can analyze chest X-rays & CT scans for COVID-19 detection.

• Comparable performance: CT scans have 98% sensitivity, X-rays 69%, making them viable diagnostic tools.

• Existing gaps: Current systems lack automated classification & severity tracking.

Market Demand

• COVID-19 diagnostics remains a priority: Rapid, accurate detection is essential for outbreak management.

• AI adoption in healthcare is growing: Increasing investment in automation for diagnostics & triage.

• Clinician need for AI support: Faster imaging-based tools can assist with triage while waiting for RT-PCR results.

Strategic Fit

• Bridges a diagnostic gap: Faster and more accessible alternative to slow lab-based testing.

• Scalable beyond COVID-19: Potential for pneumonia detection & broader AI-driven radiology applications.

Opportunity

An AI-powered COVID-19 diagnostic tool can enhance screening speed, reduce reliance on RT-PCR, and improve decision-making for healthcare professionals.

3. Prototyping the AI-Powered COVID-19 Detection System

Building & Evaluating Machine Learning Models

To develop an AI-powered COVID-19 detection system, we focused on:

• Preprocessing chest X-ray datasets for training deep learning models.

• Prototyping & optimizing object detection models (YOLO v5, Faster R-CNN).

• Designing an analytics dashboard & mobile app prototype for usability testing.

Machine Learning Development & Model Performance

Data Collection & Preprocessing

• Sourced chest X-ray datasets from medical repositories.

• Resized images (512x512, 256x256 pixels) to optimize training.

• Applied data augmentation to improve generalization.

Model Training & Evaluation

Two deep learning architectures were tested:

1. YOLO v5: Optimized for real-time detection.

2. Faster R-CNN: Enhanced lesion segmentation.

Model Performance & Insights

• YOLO v5 Model 5 (512x512 resolution) was the optimal model, achieving:

  • Precision: 0.5776 (strong performance in identifying COVID-19 positive cases).

  • Mean Average Precision (mAP@0.5): 0.4543, indicating the model successfully identified affected lung regions in over half of the images.

• Hyperparameter Optimization:

  • Trained models for 50 epochs (Run 6 extended to 75 epochs for minimizing underfitting).

  • Adjusted learning rate, momentum, and weight decay to improve convergence & reduce overfitting.

Key Takeaway:

• The optimal YOLO v5 model can assist radiologists as a decision-support tool, identifying lung abnormalities efficiently within a human-in-the-loop workflow.

Application: Conceptualizing a Doctor-Facing Analytics Tool

Beyond model development, we explored how AI could be integrated into clinical workflows through an interactive dashboard & mobile app prototype:

• Real-time X-ray analysis & severity scoring to assist radiologists.

• Trend visualization & predictive insights for patient monitoring.

• User-friendly mobile interface for rapid screening in remote/hospital settings.

This phase validated AI’s feasibility for COVID-19 detection, laying the groundwork for further development, regulatory approval, and clinical integration.

4. Roadmap & Implementation Strategy

To structure development and deployment, I aligned the roadmap with a phased approach, prioritizing technical feasibility, clinical validation, and adoption.

Phase 1: Data Collection & Model Prototyping (Completed)

• Objective: Establish technical feasibility through dataset preparation and AI model development.

  • Collected & Preprocessed X-ray Data: Standardized imaging formats for training.

  • Trained YOLO v5 & Faster R-CNN Models: Iterated on hyperparameters for optimal COVID-19 detection.

  • Validated Model Performance: Identified YOLO v5 Run 5 as the best-performing model for clinical applications.

Phase 2: Model Testing & UX Design (Completed)

• Objective: Validate model accuracy and conceptualize user-facing solutions.

  • Tested Models on Various Resolutions: Assessed trade-offs in accuracy & processing speed.

  • Developed Model Interpretability Tools: Enabled transparency in AI-driven decisions.

  • Designed AI-Powered Dashboard & Mobile Prototype: Conceptualized doctor-facing UI for real-time analysis.

Phase 3: Clinical Validation & Feedback Iteration (Future Work)

• Objective: Ensure real-world clinical adoption through testing and compliance.

  • Pilot Testing with Radiologists: Evaluate AI-assisted workflows in hospitals.

  • Refine Model with Clinician Feedback: Improve accuracy & usability based on real-world testing.

  • Address Regulatory & Compliance Needs: Align with FDA & HIPAA standards for AI-driven diagnostics.

Phase 4: Deployment & Scalability (Future Work)

• Objective: Expand adoption and integrate AI into broader radiology applications.

  • Expand to Additional Imaging Modalities: Adapt AI for CT scans & lung ultrasounds.

  • Develop Full-Scale SaaS Platform: Integrate AI-powered diagnostics into hospital IT systems.

  • Enable Continuous Model Learning: Implement real-world data feedback loops for continuous improvement.

Key Takeaways: How This Showcases My Product Thinking

AI-Driven Feasibility Testing

Validated model performance across multiple architectures & hyperparameters.

Human-Centered Design

Designed UX prototypes tailored for radiologists & frontline healthcare workers.

Scalable Roadmap

Defined phased implementation strategy for AI integration in real-world clinical settings.