AI IN HEALTHCARE
Entropy-Based Convolutional Layer Selection for Transfer Learning in Brain Tumor Detection
Deep learning is transforming medical diagnosis, but training complex models for rare conditions like brain tumors is challenging. This paper introduces an Entropy-Based Fine-Tuning method that intelligently selects which layers of a pre-trained network to retrain, significantly reducing computational demands while maintaining high accuracy for brain tumor classification.
Executive Impact & Key Findings
Our Entropy-Based Fine-Tuning approach delivers significant efficiency gains and superior diagnostic accuracy, making advanced AI more accessible for critical healthcare applications.
0%
ResNet18 Parameter Reduction
0%
ResNet34 Parameter Reduction
0%
Peak Classification Accuracy
0%
Performance Improvement vs. Prior Work
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Intelligent Layer Selection for Optimal Transfer Learning
Our novel Entropy-Based Fine-Tuning method redefines transfer learning by moving beyond traditional full or final-layer retraining. We leverage Shannon Entropy to quantify information richness in convolutional layers. The core insight is that layers with lower entropy are often underutilized and thus ideal candidates for fine-tuning. This targeted approach significantly reduces the number of trainable parameters by approximately 10% for ResNet18 and 18% for ResNet34, leading to substantial computational savings without compromising classification performance.
Accelerating Diagnosis with Pre-trained Models
Transfer Learning (TL) is crucial for medical imaging tasks like brain tumor classification, where large datasets are often scarce. Instead of training complex CNNs from scratch, TL adapts pre-trained networks (e.g., ResNet18, ResNet34) from broader tasks to specialized medical problems. Traditional TL often involves retraining only the final fully connected layers. Our research advances this by proposing a data-driven method for selective retraining of convolutional layers, ensuring more efficient and effective adaptation.
Validating Performance Across Diverse Conditions
Extensive ablation studies confirm the robustness and generalizability of our Entropy-Based Fine-Tuning method. We demonstrated its resilience to noise perturbations (Gaussian noise at various levels: 0.1, 0.2, 0.3, 0.4) and random layer selection, consistently matching or surpassing full fine-tuning performance. Furthermore, the method maintained competitive performance even when applied to a lower-capacity network, SqueezeNetV1, on a mixed dataset with additional transforms, proving its adaptability to constrained environments.
Enterprise Process Flow
Feed Images to Pre-trained Network
→
Calculate Convolutional Layer-wise Entropy
→
Identify Low-Entropy Layers
→
Unfreeze Selected Layers for Training
→
Fine-tune Network with New Task
18%
Reduction in Trainable Parameters (ResNet34)
| Method | Key Advantage | Accuracy |
|---|---|---|
| Aamir et al. [40] | Custom CNN | 97.00% |
| Amarnath et al. [41] | Xception with Modified Top Layer | 98.17% |
| Our Work (ResNet18) | Entropy-Based Layer Selection, Reduced Params | 99.05% |
| Our Work (ResNet34) | Entropy-Based Layer Selection, Reduced Params | 99.29% |
Case Study: Brain Tumor Classification with Entropy-Based Fine-Tuning
In a critical application for healthcare, our Entropy-Based Fine-Tuning method was applied to brain tumor classification using ResNet18 and ResNet34 on two diverse datasets (Kaggle and Brisc2025). The approach not only matched the high accuracy of full fine-tuning (achieving up to 99.29% accuracy) but also drastically reduced the computational resources required. This efficiency, combined with robust performance against noise and randomization, makes it an ideal candidate for rapid and accurate diagnosis in clinical settings, potentially saving critical time and resources in patient care.
Calculate Your Potential AI ROI
Estimate the efficiency gains and cost savings your enterprise could achieve by implementing smart AI solutions.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A phased approach to integrate cutting-edge AI, ensuring seamless adoption and measurable success within your enterprise.
Phase 01: Strategic Assessment & Planning
Identify high-impact areas for AI integration, define success metrics, and lay the architectural groundwork.
Phase 02: Pilot Program & Prototype Development
Develop and test initial AI solutions on a smaller scale, gathering critical feedback and refining models.
Phase 03: Scaled Deployment & Integration
Roll out AI solutions across the enterprise, integrating with existing systems and workflows.
Phase 04: Performance Monitoring & Optimization
Continuously track AI model performance, update data pipelines, and iterate for sustained improvement.
Ready to Transform Your Enterprise with AI?
Schedule a personalized consultation with our AI experts to explore how these insights can be tailored to your business challenges and drive real-world impact.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat