Out of Domain Generalization in Medical Imaging via Vision Language Models

Team

• E/19/094, Eashwara M., email
• E/19/372, Silva A.K.M, email
• E/19/408, Ubayasiri S.J, email

Supervisors

• Dr. Damayanthi Herath, email
• Dr. Ruwan Tennakoon, email

Table of content

1. Abstract
2. Introduction and Background
3. Problem Statement
4. Related Works
5. Methodology
6. BiomedXPro Architecture
7. Experimental Setup and Implementation
8. Results and Analysis
9. Additional Experiments
10. Limitations
11. Future Directions
12. Impact and Contributions
13. Conclusion
14. Publications
15. Links

Abstract

This research addresses critical domain generalization challenges in medical imaging by introducing BiomedXPro, a novel framework that leverages Vision-Language Models (VLMs) with interpretable prompt optimization. Using BiomedCLIP as the baseline model, we propose an evolutionary algorithm-based automatic prompting method that significantly improves both interpretability and out-of-domain generalization through iterative feedback mechanisms with Large Language Models (LLMs).

Our approach specifically targets disease classification tasks from histopathological images, addressing the fundamental challenge of maintaining model performance when deploying AI systems across different medical institutions with varying equipment, protocols, and patient demographics. The framework generates human-readable diagnostic prompts that capture clinically relevant visual discriminative features, ensuring both robust performance and explainable decision-making processes essential for clinical adoption.

Key achievements include achieving 93.06% accuracy on the CAMELYON17 dataset while maintaining full interpretability, demonstrating superior out-of-domain generalization across multiple hospital centers, and providing quantifiable contributions of each diagnostic observation to final predictions.

Introduction and Background

Medical Image Analysis Challenges

Medical image analysis plays a crucial role in modern healthcare, encompassing disease diagnosis, treatment guidance, image segmentation, and various other clinical applications. While significant advances have been made in medical AI systems, several critical challenges persist when deploying these systems in real-world clinical scenarios:

Domain Shift Problem

Domain shift represents one of the most significant barriers to successful deployment of AI systems in medical imaging. This phenomenon occurs when there are discrepancies between the training distribution (source domain) and the unseen distribution (target domain) where the model is deployed. In medical imaging contexts, domain shift manifests through:

• Cross-Site Variability: Different hospitals use varying equipment manufacturers, imaging protocols, and acquisition parameters
• Equipment Differences: Variations in scanner specifications, software versions, and calibration procedures
• Institutional Protocols: Different preprocessing techniques, image enhancement methods, and quality control standards
• Patient Demographics: Variations in population characteristics, disease prevalence, and comorbidities across different geographic regions

The consequence of domain shift is significant performance degradation when models trained on data from one institution are deployed in new clinical settings, potentially compromising diagnostic accuracy and patient safety.

Explainability Requirements

Modern healthcare demands AI systems that not only perform accurately but also provide transparent, interpretable decision-making processes. Medical professionals require:

• Clinical Reasoning: Understanding why a particular diagnosis was made
• Confidence Assessment: Quantifiable measures of diagnostic certainty
• Feature Attribution: Identification of specific visual features that contributed to the decision
• Regulatory Compliance: Meeting standards for medical device approval and clinical validation

Vision Language Models in Medical Imaging

Vision Language Models (VLMs), particularly Contrastive Language-Image Pretraining (CLIP) and its biomedical variant BiomedCLIP, have emerged as promising solutions for addressing both domain generalization and explainability challenges:

Key Advantages of VLMs:

1. Zero-Shot Classification Capabilities: Models can classify images without task-specific training by leveraging natural language descriptions
2. Inherent Interpretability: Classifications are based on natural language prompts that humans can understand and validate
3. Robustness to Distribution Shifts: Pre-training on diverse datasets provides inherent resilience to domain variations
4. Multimodal Understanding: Joint representation learning enables sophisticated reasoning about visual-textual relationships

BiomedCLIP Architecture

BiomedCLIP represents the state-of-the-art in biomedical vision-language models, specifically designed for medical imaging applications. The model architecture consists of:

• Vision Encoder: Processes medical images and extracts visual features
• Text Encoder: Processes natural language descriptions and creates textual representations
• Shared Embedding Space: Aligns visual and textual representations for similarity-based classification
• Contrastive Learning: Optimizes the model to maximize similarity between matching image-text pairs while minimizing similarity between non-matching pairs

Problem Statement

Despite the promising capabilities of biomedical CLIP models, current approaches face a critical limitation that hinders their clinical adoption:

Core Challenge: Biomedical CLIP models demonstrate inherent robustness to distribution shifts, and their performance can be significantly enhanced through context optimization techniques. However, existing context optimization methods rely on uninterpretable “soft” prompts - learned vector representations that lack human readability.

Clinical Implications: This lack of interpretability presents a fundamental barrier for clinical adoption, where both out-of-domain generalization AND explainability are paramount requirements for reliable AI-driven diagnostics. Medical professionals cannot trust or validate diagnostic decisions based on abstract vector representations that provide no clinical reasoning.

Specific Limitations of Current Approaches:

1. Soft Vector Learning: Context optimization methods like CoOp generate numerical vectors (e.g., [1.3, 2.3, 4.2, …]) that cannot be interpreted by medical professionals
2. Single Static Outputs: Many existing methods rely on single LLM outputs without iterative refinement
3. Limited Clinical Validation: Generated prompts often lack verification against established medical knowledge
4. Insufficient Diversity: Single optimal prompt approaches fail to capture the complexity of medical diagnostic reasoning

Related Works

Vision-Language Models in Biomedical Applications

The development of vision-language models has revolutionized biomedical image analysis, with several key contributions shaping the field:

Foundation Models

CLIP (Contrastive Language-Image Pretraining): Introduced the concept of learning joint representations of images and text through contrastive learning. While effective for natural images, CLIP’s performance on specialized biomedical images remained limited due to domain-specific terminology and visual characteristics.

BiomedCLIP: Specifically designed for biomedical applications, this model was trained on biomedical image-text pairs, significantly improving performance on medical imaging tasks while maintaining the interpretability advantages of the original CLIP architecture.

Prompt Learning Approaches

BiomedCoOp: Extended the Context Optimization (CoOp) approach to biomedical domains, learning continuous prompt vectors that optimize classification performance. However, these learned vectors lack interpretability, making clinical validation challenging.

XCoOp: Introduced cross-modal prompt learning, attempting to bridge vision and language modalities more effectively. Despite improved performance, the fundamental interpretability limitation persisted.

Limitations of Existing Approaches

1. Interpretability Gap: Most existing methods focus solely on performance optimization, neglecting the critical need for explainable diagnostic reasoning
2. Single Prompt Limitation: Many approaches optimize for a single “best” prompt, failing to capture the diversity of diagnostic observations
3. Static Generation: Limited use of iterative refinement processes that could improve prompt quality over time
4. Clinical Validation Deficit: Insufficient integration of medical domain expertise in prompt generation and validation

Gap in Current Research

Our comprehensive literature review revealed a significant gap: no existing method successfully combines high performance with interpretable prompt generation for biomedical vision-language models. This gap represents a critical barrier to clinical adoption and motivated our development of BiomedXPro.

Methodology

Our methodology introduces BiomedXPro, a novel framework that addresses the interpretability-performance trade-off through evolutionary prompt optimization. The approach consists of several integrated components:

1. Preprocessing Pipeline

Data Quality Assurance

• File Integrity Checks: Verification of image file completeness and format consistency
• Metadata Validation: Ensuring all required clinical annotations are present and properly formatted
• Quality Control: Identification and handling of corrupted or low-quality images

Image Standardization

• Normalization: Standardizing pixel values across different acquisition protocols and equipment
• Resizing: Uniform resizing to 224×224 pixels to match BiomedCLIP input requirements
• Color Space Consistency: Ensuring consistent color representation across different imaging sources

Dataset Preparation

• Label Unification: Standardizing diagnostic labels across different institutional coding systems
• Demographic Balancing: Ensuring representative distribution of patient demographics and disease subtypes
• Domain-Aware Splitting: Creating training/validation/test splits that respect domain boundaries for proper out-of-domain evaluation

2. Evolutionary Prompt Optimization Framework

Theoretical Foundation

Our approach leverages Large Language Models as implicit optimizers, drawing inspiration from evolutionary algorithms and gradient-free optimization techniques. The key insight is that LLMs can serve multiple roles:

• Prompt Engineers: Generating diverse, clinically relevant diagnostic descriptions
• Knowledge Extractors: Leveraging vast medical knowledge to create meaningful prompts
• Optimizers: Using performance feedback to iteratively improve prompt quality

LLM Integration Strategy

We utilize Gemma3 27B as our primary LLM, selected for its strong performance in medical domain tasks and cost-effectiveness for iterative optimization processes.

3. Multi-Stage Optimization Process

Stage 1: Initial Prompt Population Generation

Meta-Prompt Design (Q₀):

Give 50 textual description pairs of visual discriminative features to identify whether the central region of a histopathological image patch contains tumor tissue or not. The patch is extracted from an H&E‑stained whole‑slide image of a lymph node section.

This initial meta-prompt is carefully crafted to:

• Specify the exact clinical context (histopathological lymph node analysis)
• Request discriminative feature pairs (positive and negative classes)
• Generate sufficient diversity (50 initial prompts) for robust optimization

Stage 2: Fitness Evaluation and Selection

Fitness Score Calculation: The fitness function evaluates each prompt pair based on classification performance:

f(p) = Performance_Metric(BiomedCLIP_with_prompt_p, training_data)

We experimented with multiple fitness metrics:

• Accuracy: Overall classification correctness
• AUC (Area Under Curve): Discrimination capability across thresholds
• Inverted Cross-Entropy Loss: Focusing on confident correct predictions
• F1 Score: Balanced consideration of precision and recall

Roulette Wheel Selection: This probabilistic selection mechanism ensures:

• Exploitation: Higher-performing prompts have greater selection probability
• Exploration: Lower-performing prompts retain selection chance for diversity
• Balance: Prevents premature convergence to local optima

Stage 3: Iterative Prompt Generation

Optimizer Meta-Prompt (Qᵢ):

The task is to generate textual descriptions pairs of visual discriminative features to identify whether the central region of an histopathological image patch contains tumor tissue or not.

Here are the best performing pairs in descending order:
1. (..., ...) Score: 90
2. (..., ...) Score: 84
...

Write 10 new prompt pairs that are different from the old ones and has a score as high as possible.

This iterative process enables:

• Guided Evolution: Using performance feedback to direct prompt generation
• Diversity Maintenance: Explicit instruction to generate different prompts
• Quality Improvement: Targeting higher performance scores

Stage 4: Diversity Maintenance Through Crowding

Crowding Meta-Prompt (Qc): The crowding mechanism prevents convergence to semantically identical prompts with different linguistic expressions:

Group the prompt pairs that have exactly the same medical observation but differ only in language variations. Provide the output as grouped indices.

This process:

• Identifies Redundancy: Groups semantically equivalent prompts
• Maintains Diversity: Eliminates linguistic variations while preserving conceptual diversity
• Improves Efficiency: Reduces computational overhead from redundant evaluations

4. Final Prompt Selection and Ensemble

Elbow Analysis for Optimal Prompt Count

Rather than using a fixed number of final prompts, we employ elbow analysis on the fitness score distribution to automatically determine the optimal number of prompts that maximize diversity while maintaining high performance.

Weighted Ensemble Voting

Final Classification Process:

1. Individual Prompt Evaluation: Each selected prompt pair provides a binary vote (tumor/normal)
2. Weight Calculation: Fitness scores are normalized to create prompt-specific weights
3. Weighted Aggregation: Final decision combines all votes using calculated weights
4. Threshold Application: Scores > 0.5 indicate tumor presence

Mathematical Formulation:

Final_Score = Σᵢ (Normalized_Fitness_Score_i × Vote_i)
Prediction = Final_Score > 0.5 ? "Tumor" : "Normal"

BiomedXPro Architecture

System Architecture Overview

BiomedXPro implements a sophisticated evolutionary algorithm specifically designed for medical prompt optimization. The architecture consists of several interconnected components working in harmony to achieve optimal performance while maintaining interpretability.

Core Components

1. LLM-Driven Prompt Generation Engine

• Model: Gemma3 27B for cost-effective, high-quality prompt generation
• Role: Serves as both prompt engineer and knowledge extractor
• Capability: Generates clinically relevant diagnostic descriptions based on medical domain knowledge

2. Fitness Evaluation System

• Integration: Direct connection with BiomedCLIP for real-time performance assessment
• Metrics: Multiple evaluation criteria (accuracy, AUC, F1-score, inverted BCE)
• Efficiency: Batch processing for optimal computational resource utilization

3. Evolutionary Optimization Framework

• Selection: Roulette wheel selection for balanced exploration-exploitation
• Generation: Iterative prompt creation with performance-based guidance
• Diversity: Crowding mechanisms to prevent semantic redundancy

4. Interpretability Layer

• Human-Readable Outputs: All final prompts remain in natural language form
• Clinical Validation: Generated prompts align with established medical diagnostic criteria
• Explainable Decisions: Each diagnostic decision can be traced back to specific visual features

Optimization Parameters

• Iterations: 1,000 evolutionary cycles for thorough optimization
• Priority Queue Capacity: 1,000 prompts for comprehensive exploration
• Filter Threshold: 0.6 minimum fitness score for quality control
• Initial Population: 50 diverse prompt pairs for robust starting diversity
• Generation Rate: 10 new prompts per iteration for steady evolution

Performance Characteristics

• Convergence: Typically achieves optimal performance within 800-1000 iterations
• Stability: Consistent results across multiple optimization runs
• Scalability: Framework designed to handle multiple diseases and imaging modalities
• Efficiency: Gradient-free optimization reduces computational requirements compared to traditional fine-tuning

Experimental Setup and Implementation

Dataset Configuration

Primary Dataset: CAMELYON17 WILDS

Dataset Characteristics:

• Source: Histopathological lymph node sections from multiple hospitals
• Task: Binary classification (normal tissue vs. tumor tissue)
• Domain Structure: 3 different hospitals representing distinct domains
• Scale: Large-scale dataset with thousands of high-resolution histopathological images
• Ground Truth: Expert pathologist annotations for reliable evaluation

Domain Distribution:

• Hospital 0: In-distribution training data with specific imaging protocols
• Hospital 1: In-distribution data with different acquisition parameters
• Hospital 2: In-distribution data representing third institutional variation
• Main Test Set: Out-of-domain evaluation across all hospital sites

Secondary Datasets for Generalization Testing

NIH ChestX-ray14:

• Modality: Chest X-ray radiographs
• Task: Multi-label classification with 14 disease categories
• Scale: Large-scale dataset for thoracic disease detection
• Purpose: Evaluating cross-modality generalization capabilities

CheXpert:

• Modality: Chest X-ray imaging
• Task: Multi-label classification with 14 clinical observations
• Unique Features: Uncertainty labels for realistic clinical scenarios
• Validation: Cross-dataset evaluation for chest imaging robustness

RETOUCH:

• Modality: Optical Coherence Tomography (OCT)
• Task: Detection of retinal fluid presence (3 categories)
• Specialty: Ophthalmological imaging for retinal disease assessment
• Purpose: Demonstrating applicability across medical specialties

Implementation Details

Hardware Configuration

• GPU: High-performance computing environment for BiomedCLIP inference
• Memory: Sufficient RAM for large-scale image batch processing
• Storage: High-speed storage for efficient dataset loading and caching

Software Framework

• Base Model: BiomedCLIP for vision-language understanding
• LLM Integration: Gemma3 27B via API for prompt generation
• Optimization: Custom evolutionary algorithm implementation
• Evaluation: Comprehensive metrics calculation and statistical analysis

Hyperparameter Configuration

Evolutionary Algorithm Parameters:

• Population Size: 50 initial prompts for adequate genetic diversity
• Selection Pressure: Roulette wheel with fitness-proportionate selection
• Mutation Rate: Controlled through LLM generation diversity parameters
• Elitism: Top performers preserved across generations
• Convergence Criteria: Performance plateau detection over 100 iterations

LLM Generation Parameters:

• Temperature: Optimized for balance between creativity and coherence
• Top-p Sampling: Nucleus sampling for high-quality prompt generation
• Max Tokens: Sufficient length for detailed clinical descriptions
• Safety Filtering: Medical domain appropriateness validation

Evaluation Methodology

Performance Metrics

• Primary Metrics: Accuracy, AUC, F1-score for classification performance
• Secondary Metrics: Precision, recall, sensitivity, specificity for clinical relevance
• Interpretability Metrics: Prompt coherence, clinical validity, expert assessment

Cross-Validation Strategy

• Domain-Aware Splitting: Ensuring proper separation of training and test domains
• Stratified Sampling: Maintaining class balance across all evaluation splits
• Multiple Runs: Statistical significance testing across multiple optimization runs

Baseline Comparisons

• Zero-Shot BiomedCLIP: Original model with manually engineered prompts
• BiomedCLIP + CoOp: Context optimization with learned soft vectors
• Standard Fine-Tuning: Traditional supervised learning approaches
• Existing Prompt Learning Methods: Comparison with state-of-the-art techniques

Results and Analysis

CAMELYON17 Primary Results

Evolutionary Optimization Progress

The optimization process demonstrates clear convergence characteristics over 1,000 iterations:

Convergence Analysis:

• Initial Performance: Starting accuracy around 75-80% with basic prompt generation
• Rapid Improvement Phase: Significant gains in first 200-300 iterations as LLM identifies effective diagnostic features
• Refinement Phase: Gradual improvement from iterations 300-800 as prompts become more clinically precise
• Convergence: Performance plateau achieved around iteration 800-900, indicating optimal prompt discovery

Final Optimized Prompts (Top 8 selected via elbow analysis):

1. Primary Diagnostic Prompt (Score: 0.9013):
   • Negative: “No atypical cells infiltrating surrounding tissues”
   • Positive: “Atypical cells infiltrating surrounding tissues and disrupting normal structures”
2. Cellular Atypia Assessment (Score: 0.8997):
   • Negative: “No significant atypia in the surrounding lymphocytes”
   • Positive: “Significant atypia observed in lymphocytes adjacent to tumor nests”
3. Stromal Changes Detection (Score: 0.8994):
   • Negative: “No evidence of fibrosis”
   • Positive: “Prominent stromal fibrosis surrounding tumor nests”
4. Architectural Preservation (Score: 0.8940):
   • Negative: “Normal follicular architecture is preserved”
   • Positive: “Disrupted follicular architecture with loss of polarity”

These prompts demonstrate sophisticated understanding of histopathological features that pathologists use for tumor diagnosis, including cellular infiltration patterns, nuclear atypia, stromal reactions, and architectural disruption.

Comparative Performance Analysis

Key Performance Insights

Competitive Performance: BiomedXPro achieves 93.06% accuracy on the main test set, demonstrating only a 0.84% performance gap compared to CoOp while providing full interpretability.

Domain Generalization: Strong performance across all hospital domains indicates robust out-of-domain generalization, with particularly impressive results on Hospital 2 (93.69%).

Interpretability Advantage: Unlike CoOp’s uninterpretable context vectors, BiomedXPro provides clinically meaningful prompts that medical professionals can understand and validate.

Detailed Method Comparison

BiomedXPro vs. CoOp Analysis

Statistical Significance Analysis

Performance Distribution: Multiple optimization runs (n=5) show consistent results:

• Mean Accuracy: 93.01% ± 0.12%
• Standard Deviation: Low variance indicating stable optimization
• Confidence Interval: 95% CI [92.89%, 93.13%]

Domain Robustness: Cross-domain evaluation demonstrates:

• Consistent Performance: Less than 4% accuracy drop across domains
• Balanced Results: No single domain showing exceptional vulnerability
• Generalization Capability: Strong performance on completely unseen data

Additional Experiments

Multi-Dataset Evaluation

Cross-Modality Generalization

NIH ChestX-ray14 Results:

• Performance: Successfully adapted prompts for chest X-ray analysis
• Challenge: Multi-label classification requiring simultaneous disease detection
• Outcome: Demonstrated framework flexibility across imaging modalities

CheXpert Validation:

• Cross-Dataset Evaluation: Model trained on NIH data tested on CheXpert
• Robustness: Maintained performance across different chest X-ray datasets
• Clinical Relevance: Generated prompts aligned with radiological reporting standards

RETOUCH OCT Analysis:

• Specialty Adaptation: Extended framework to ophthalmological imaging
• Task Complexity: Retinal fluid detection in OCT images
• Success: Effective prompt generation for specialized medical imaging

LLM Comparison Study

Performance Variation Across Different LLMs

ChatGPT 4.1 vs. Gemma3 27B:

• Quality Assessment: Both models generated clinically relevant prompts
• Cost Analysis: Gemma3 27B provided better cost-effectiveness for iterative optimization
• Performance Differences: Marginal variations in final classification accuracy
• Consistency: Both showed stable convergence characteristics

Key Findings:

• Model Choice Flexibility: Framework successfully adapts to different LLM capabilities
• Cost-Performance Trade-off: Gemma3 27B offers optimal balance for production deployment
• Quality Maintenance: High-quality prompt generation maintained across model choices

Advanced Ensemble Methods

Stacking Approach Evaluation

Meta-Model Performance on CAMELYON17:

Stacking Benefits:

• Relationship Capture: Meta-models learn complex relationships between prompt observations
• Performance Enhancement: Marginal improvements over simple weighted voting
• Robustness: Reduced variance in predictions through ensemble diversity

Implementation Considerations:

• Additional Complexity: Requires training meta-model on prompt outputs
• Computational Overhead: Increased inference time for meta-model predictions
• Interpretability Impact: Additional layer reduces direct prompt-to-decision traceability

Ablation Studies

Component Contribution Analysis

Evolutionary Components:

1. Roulette Wheel Selection: 2.3% performance improvement over random selection
2. Crowding Mechanism: 1.8% improvement through diversity maintenance
3. Iterative Refinement: 4.5% improvement over single-generation prompts
4. Ensemble Voting: 1.2% improvement over single best prompt

Optimization Parameters:

• Population Size Impact: Optimal performance at 50 initial prompts
• Iteration Requirements: Convergence typically achieved by iteration 800
• Selection Pressure: Balanced exploration-exploitation at current settings

Limitations

Clinical Validation Requirements

Expert Feedback Integration

Pathologist Review (Dr. Sumanarasekara):

• Prompt Specificity: Some generated prompts were identified as too vague for precise diagnostic use
• Tumor Type Generalization: Generated prompts must accommodate different tumor subtypes within lymph node metastases
• Clinical Context: Need for more specific anatomical and morphological descriptors

Identified Issues:

1. Generality vs. Specificity Trade-off: Balancing broad applicability with diagnostic precision
2. Medical Terminology: Ensuring proper use of standardized pathological terminology
3. Diagnostic Hierarchy: Incorporating proper diagnostic decision trees used by pathologists

Validation Framework Needs

Expert Integration Process:

• Systematic Review: Structured evaluation protocol for medical expert assessment
• Feedback Incorporation: Mechanism to integrate expert corrections into optimization process
• Continuous Improvement: Ongoing refinement based on clinical usage feedback

Computational and Economic Constraints

LLM Generation Costs

Current Cost Structure:

• Iterative Process: 1,000 iterations with 10 prompts each generation
• API Costs: Significant expenses for large-scale optimization using commercial LLMs
• Scalability Challenge: Cost multiplication when extending to multiple diseases/modalities

Cost Optimization Strategies:

• Local Model Deployment: Using open-source models for reduced API costs
• Batch Processing: Optimizing generation requests for improved efficiency
• Caching Mechanisms: Reusing successful prompts across similar tasks

Performance Trade-offs

BiomedXPro vs. CoOp Performance Gap:

• Accuracy Difference: 0.84% lower performance compared to CoOp
• Interpretability Benefit: Significant gain in explainability at minimal performance cost
• Clinical Acceptance: Trade-off favors interpretability for clinical adoption

Scalability Considerations

Multi-Disease Extension

Current Limitations:

• Single Disease Focus: Primary validation on lymph node metastasis detection
• Prompt Transferability: Uncertainty about prompt reuse across different pathologies
• Optimization Overhead: Separate optimization potentially required for each new disease

Required Developments:

• Multi-Disease Framework: Simultaneous optimization across multiple pathological conditions
• Transfer Learning: Leveraging successful prompts for related diagnostic tasks
• Hierarchical Organization: Structuring prompts by organ system and pathology type

Dataset Dependency

Training Data Requirements:

• Domain Representation: Need for adequate representation of all target domains
• Quality Standards: Dependence on high-quality expert annotations
• Bias Mitigation: Ensuring diverse demographic and institutional representation

Conclusion

The integration of iterative, interpretable prompt generation using LLMs significantly improves the domain generalization capabilities of vision-language models in medical imaging. The approach offers a path forward for deploying robust and explainable AI tools in clinical settings.

Publications

Links

• Project Repository
• Project Page
• Department of Computer Engineering
• University of Peradeniya