A Centralized AI Lakehouse Framework for Brain Tumor MRI Classification and Segmentation, University KPI Forecasting,...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Engineering

Background:

• University and healthcare projects often use disparate AI models for varied data types (tables, time series, images).
• Separate model deployments lead to testing and maintenance challenges due to unique pipelines and runtime requirements.

Purpose of the Study:

• To present an integrated AI lakehouse-style implementation for deploying multiple model pipelines within a single containerized backend.
• To demonstrate the framework's versatility across medical imaging, university KPI forecasting, and water potability classification tasks.

Main Methods:

• Developed a containerized backend using Docker and FastAPI to host three distinct AI model pipelines.
• Utilized ResNet50 for medical image classification (99.00% accuracy) and DeepLabV3-ResNet50 for segmentation (83.09% mIoU).
• Applied Random Forest and ARIMA for university KPI forecasting and Random Forest for water potability classification (85.86% accuracy).

Main Results:

• Achieved high performance in medical image classification and segmentation tasks.
• Demonstrated realistic university KPI forecasting by evaluating models on original annual data, not interpolated data.
• Validated the Random Forest model for water potability with 85.86% accuracy and high recall for unsafe samples.
• System-level benchmarking confirmed stable throughput and latency under concurrent requests.

Conclusions:

• The integrated AI lakehouse approach simplifies deployment and maintenance of diverse AI models.
• This unified framework offers a scalable and efficient solution for complex data science projects in healthcare and academia.
• The system effectively handles multiple model types and data modalities within a single, robust backend.