Image Classification and ML in Disease Recognition: A Research Review

# Image Classification and ML in Disease Recognition: A Research Review

**Medical ML Research Series**

**By Oleh Ivchenko, PhD Candidate**
**Affiliation:** Odessa Polytechnic National University | Stabilarity Hub | February 2026

—

94.4%

Skin Cancer Detection Accuracy

50-60%

Error Reduction with Human-AI

140

Radiologists in Nature Med Study

Pipeline Stages with ML Integration

—

## Introduction

**Medical image analysis stands at a transformative crossroads.** As deep learning models achieve remarkable accuracy in disease detection, a critical question emerges: how do we integrate AI into clinical workflows to maximize diagnostic accuracy while minimizing errors? This comprehensive review examines the current state of ML in medical imaging, mapping which techniques apply at each diagnostic stage, and synthesizing evidence-based best practices for human-AI collaboration.

—

## The Medical Image Analysis Pipeline

graph LR A[Preprocessing] --> B[Segmentation] B --> C[Features] C --> D[Classification] D --> E[Explainability]

Stage	ML Technique	Primary Models	Clinical Impact
1. Preprocessing	Image enhancement, noise reduction	Autoencoders, GANs	Improved image quality
2. Segmentation	ROI detection	U-Net, SAM	Automated lesion boundaries
3. Feature Extraction	Deep feature learning	ResNet, VGG, DenseNet	Capture patterns
4. Classification	Disease prediction	CNN, ViT, Hybrid	Malignancy scoring
5. Explainability	Decision interpretation	Grad-CAM, SHAP	Physician understanding

—

## ML Model Evolution

graph TD A[Traditional CNNs] --> B[Transfer Learning] B --> C[Vision Transformers] C --> D[Hybrid Architectures]

—

## Case Studies: Performance Metrics

🔬 Skin Cancer Detection

94.4%

EViT-DenseNet169 on HAM10000

Hybrid attention + multi-scale fusion

🔬 Breast Cancer

92%

Self-Attention CNN + GA

Inverted residual blocks + feature selection

🔬 Diabetic Retinopathy

97%

Vision Transformer

Global context attention

—

## Doctor-AI Collaboration Framework

graph LR A[Medical Image] --> B[AI Analysis] B --> C[Doctor Review] C --> D[Final Diagnosis]

Factor	Doctor Strength	AI Strength	Combined Advantage
Clinical Context	Patient history, symptoms	Pure image analysis	Contextualized interpretation
Consistency	Subject to fatigue	Consistent 24/7	AI catches fatigue errors
Speed	Minutes per case	Seconds per case	AI pre-screens, doctor reviews
Accountability	Legal responsibility	No liability	Doctor maintains authority

—

## Tiered Review Protocol

graph TD A[AI Analysis] --> B[High Confidence] A --> C[Moderate Confidence] A --> D[Low Confidence] B --> E[Fast Track] C --> F[Standard Review] D --> G[Expert Review]

—

## ML Architectures by Disease Domain

Disease Domain	Modality	Best Architecture	Accuracy
Skin Cancer	Dermoscopy	EViT-DenseNet169	92-95%
Breast Cancer	Mammography	Self-Attention CNN	90-93%
Lung Nodules	CT Scan	3D ResNet + Attention	87-91%
Diabetic Retinopathy	Fundoscopy	Vision Transformer	94-97%
Colorectal Cancer	Histopathology	MIL + Foundation	89-92%

—

## Unique Conclusions

🔬 Conclusion 1: The Heterogeneity Paradox

Physician experience does not predict who benefits most from AI assistance. Universal deployment with individualized feedback is more effective than experience-based targeting.

🔬 Conclusion 2: The Hybrid Architecture Advantage

Across disease domains, hybrid CNN-Transformer architectures consistently outperform pure approaches. CNNs excel at local features while Transformers capture global context.

🔬 Conclusion 3: Adaptive Explainability

Optimal solution is adaptive explainability: detailed explanations only when AI-physician disagreement occurs or confidence is low.

—

## References

1. Ly, N. et al. “Recent Advances in Medical Image Classification.” arXiv:2506.04129, 2025.
2. Agarwal, N. et al. “Heterogeneity and predictors of AI effects on radiologists.” *Nature Medicine*, 2024.
3. Chen, R.J. et al. “A pathologist–AI collaboration framework.” *Nature Biomedical Engineering*, 2024.
4. “Enhanced early skin cancer detection through EViT-DenseNet169.” *Scientific Reports*, 2025.
5. “Artificial intelligence based classification using inverted self-attention DNN.” *Scientific Reports*, 2025.

—

**Author:** Oleh Ivchenko, PhD Candidate
**Affiliation:** Odessa Polytechnic National University | Stabilarity Hub