StageMamba: Stage-Aware Multi-Level Feature Fusion with Mamba Blocks for Multi-Class Eye Disease Diagnosis

Accepted at IHCI 2025 (17th International Conference on Intelligent Human-Computer Interaction)
Published in LNCS Springer Proceedings | Paper ID: 308
📍 Marriott Hotel, Jaipur, Rajasthan, India — November 14–16, 2025

---

Overview

StageMamba is a novel deep learning framework for automated multi-class eye disease diagnosis from retinal fundus images. It integrates:

• EfficientNet-B4 backbone for hierarchical multi-stage feature extraction
• Multi-Level Feature Fusion (MLFF) module with Stage Enhancers, Lateral Reduce, Fuse Mix, and Unify Channel operations
• Vision State Space (VSS) module with Mamba blocks for efficient long-range dependency modeling
• MLP classifier for 10-class disease prediction

StageMamba achieves 90.03% accuracy, F1 = 0.9015, and MCC = 0.8897 on the augmented dataset — surpassing all CNN baselines by +3.26% accuracy.

---

Architecture

Input Fundus Image
        │
 EfficientNet-B4 Backbone
 (Stage 1 → Stage 5 features)
        │
 Multi-Level Feature Fusion (MLFF)
 ┌──────────────────────────────┐
 │ Stage Enhancer (C2–C5)      │
 │ Lateral Reduce (1×1 Conv)   │
 │ Fuse Mix (Cross-stage)      │
 │ Unify Channel               │
 └──────────────────────────────┘
        │
 VSS-Mamba Encoder
 (Long-range dependencies, linear complexity)
        │
 MLP Classifier
 (10 disease classes)
        │
     Output

10 Disease Classes

CSC | Diabetic Retinopathy | Disc Edema | Glaucoma | Healthy | Macular Scar | Myopia | Pterygium | Retinal Detachment | Retinitis Pigmentosa

---

Results

Baseline CNN Performance on Original Dataset

Model	Accuracy	Precision	Recall	F1-Score	MCC
ResNet34	0.8012	0.8045	0.8012	0.8006	0.7751
VGG19	0.8078	0.8093	0.8078	0.8065	0.7812
ResNet50	0.8125	0.8151	0.8125	0.8119	0.7860
MobileNetV2	0.8186	0.8194	0.8186	0.8179	0.7928
DenseNet121	0.8241	0.8256	0.8241	0.8233	0.7990
InceptionV3	0.8290	0.8302	0.8290	0.8288	0.8041
EfficientNet-B4	0.8334	0.8340	0.8334	0.8331	0.8096

Baseline CNN Performance on Augmented Dataset

Model	Accuracy	Precision	Recall	F1-Score	MCC
ResNet34	0.8492	0.8498	0.8492	0.8492	0.8236
VGG19	0.8535	0.8550	0.8535	0.8532	0.8292
ResNet50	0.8541	0.8548	0.8541	0.8538	0.8296
MobileNetV2	0.8590	0.8601	0.8590	0.8591	0.8353
DenseNet121	0.8627	0.8628	0.8627	0.8626	0.8395
InceptionV3	0.8677	0.8685	0.8677	0.8672	0.8453
EfficientNet-B4	0.8677	0.8680	0.8677	0.8676	0.8456

StageMamba vs Best Baseline (Augmented)

Model	Accuracy	F1-Score	MCC
EfficientNet-B4 (Baseline)	86.77%	0.8676	0.8456
StageMamba (Proposed)	90.03%	0.9015	0.8897

Ablation Study

Variant	Accuracy	F1	MCC
EffNet-B4 (backbone only)	86.77	0.8676	0.8456
EffNet-B4 + MLFF	88.10	0.8812	0.8641
EffNet-B4 + Mamba	88.65	0.8871	0.8712
StageMamba (Full)	90.03	0.9015	0.8897

---

Dataset

Eye Disease Image Dataset — Mendeley Data V1, 2024

Class	Original	Augmented
Central Serous Chorioretinopathy	101	606
Diabetic Retinopathy	1509	3444
Disc Edema	127	762
Glaucoma	1349	2880
Healthy	1024	2676
Macular Scar	444	1937
Myopia	500	2251
Pterygium	17	102
Retinal Detachment	125	750
Retinitis Pigmentosa	139	834
Total	5,335	16,492

---

Installation

git clone https://github.com/your-username/StageMamba.git
cd StageMamba
pip install -r requirements.txt

---

Usage

1. Train Baseline CNN Models

python scripts/train_baselines.py \
    --data_dir /path/to/dataset \
    --save_dir ./results/baselines \
    --epochs 10 \
    --batch_size 32 \
    --lr 1e-4

2. Train StageMamba

python scripts/train_stagemamba.py \
    --data_dir /path/to/augmented_dataset \
    --save_dir ./results/stagemamba \
    --epochs 30 \
    --batch_size 16 \
    --lr 1e-4

3. Evaluate a Trained Model

python scripts/evaluate.py \
    --model_path ./results/stagemamba/stagemamba_best.pth \
    --data_dir /path/to/test_data \
    --model_type stagemamba

4. Run Augmentation Pipeline

python scripts/augment_dataset.py \
    --input_dir /path/to/original \
    --output_dir /path/to/augmented \
    --target_per_class 2000

---

Project Structure

StageMamba/
├── src/
│   ├── models/
│   │   ├── __init__.py
│   │   ├── stagemamba.py          # Full StageMamba model
│   │   ├── mlff.py                # Multi-Level Feature Fusion module
│   │   ├── mamba_block.py         # VSS-Mamba encoder block
│   │   └── stage_enhancer.py      # Stage Enhancer, Lateral Reduce, Fuse Mix
│   ├── data/
│   │   ├── __init__.py
│   │   ├── dataset.py             # EyeDiseaseDataset class
│   │   └── augmentation.py        # Augmentation pipeline
│   └── utils/
│       ├── __init__.py
│       ├── metrics.py             # Accuracy, F1, MCC computation
│       ├── trainer.py             # Training loop with checkpoint support
│       └── visualization.py       # Plot results, confusion matrix
├── scripts/
│   ├── train_baselines.py         # Train all baseline CNN models
│   ├── train_stagemamba.py        # Train StageMamba
│   ├── evaluate.py                # Evaluate any trained model
│   └── augment_dataset.py         # Dataset augmentation script
├── configs/
│   ├── baseline_config.yaml       # Baseline training config
│   └── stagemamba_config.yaml     # StageMamba training config
├── notebooks/
│   └── results_analysis.ipynb     # Results visualization notebook
├── results/                       # Auto-generated during training
├── docs/
│   └── architecture.md
├── requirements.txt
└── README.md