---

⚠️ REVIEW-ONLY NOTICE

This repository is provided solely for peer review and reproducibility purposes
associated with the submitted manuscript.

Reuse, redistribution, modification, or deployment of any code, data, or results
contained herein is strictly prohibited without explicit written permission from the authors.

© 2026 The Authors. All Rights Reserved.

---

---

Hierarchically Enhanced Hybrid Learning for Ichthyosis Classification(H-CoAtNet)

Official Research Codebase

This repository contains the reference implementation of H-CoAtNet, a hierarchically enhanced hybrid convolution–transformer framework for multi-class Ichthyosis subtype classification from dermatological images.

The release supports methodological verification, benchmarking, and reproducibility for rare disease medical image analysis.

---

📄 Associated Paper

**Hierarchical Hybrid Learning: Enhanced Classification of Ichthyosis Variants in Dermatological Images Using H-CoAtNet **

Athul Joe Joseph Palliparambil, Anandhu P Shaji, Rajeev Rajan (Under Review)

---

🔧 Repository Structure and Execution Context (Critical)

After cloning, note that the actual project root is the inner H-CoAtNet/ directory.

git clone https://github.com/Cyrax321/H-CoAtNet-Ichthyosis.git
cd H-CoAtNet-Ichthyosis
cd H-CoAtNet

All commands must be executed from this directory. Running commands from the outer directory will result in missing file or module errors.

H-CoAtNet/
├── README.md
├── requirements.txt
├── proposed_method/
│   └── train_h_coatnet.py
└── baselines/
    ├── train_cnn.py
    ├── train_efficientnet.py
    ├── train_vit.py
    ├── train_swin.py
    ├── train_coatnet.py
    └── train_gft.py

---

1. Environment Setup

Dependencies

pip install -r requirements.txt

Core Requirements

• Python ≥ 3.9
• PyTorch
• timm
• torchvision
• scikit-learn
• numpy, pandas, matplotlib
• roboflow

Tested on macOS (Apple Silicon) and Linux environments.

---

2. Problem Overview

Ichthyosis represents a heterogeneous group of rare genetic skin disorders characterized by abnormal keratinization and severe scaling. Automated classification is challenging due to:

• Extreme class imbalance
• Subtle morphological differences between subtypes
• Limited annotated medical datasets

H-CoAtNet addresses these challenges through hybrid convolution–transformer modeling with hierarchical feature refinement.

---

3. Method Overview

H-CoAtNet integrates three core architectural components:

• Convolutional stem for localized texture and scale modeling
• Transformer blocks for global contextual dependency learning
• Hierarchical squeeze-excitation with progressive token selection

This design balances inductive bias, global reasoning, and computational efficiency, optimized for rare disease image classification.

---

4. Dataset Description

The dataset used in this study contains 1,580 dermatological images distributed across five diagnostic categories:

• Harlequin Ichthyosis (HI)
• Ichthyosis Vulgaris (IV)
• Lamellar Ichthyosis (LI)
• Netherton Syndrome (NS)
• Healthy Skin

Images are resized to 224 × 224, normalized using ImageNet statistics, and split using stratified 70/15/15 train–validation–test partitions.

The dataset reflects clinically realistic prevalence while ensuring sufficient representation of rare subtypes.

---

5. Dataset Access and API Configuration (Required Before Running Code)

To ensure controlled access, versioning, and reproducibility, the dataset is hosted using Roboflow.

📎 Dataset Project Link

Roboflow Dataset Page https://universe.roboflow.com/hi-l9ueo/ich-s-7lnsj

---

How to Access the Data

To obtain your Roboflow API key, follow these steps:

1. Click the dataset project link above.
2. Navigate to Dataset in the left sidebar.
3. Click Download Dataset.
4. Select Download Dataset (Get a code snippet or ZIP file).
5. Ensure Show download code is enabled.
6. Choose: “Custom train this dataset using the provided code snippet in a notebook.”
7. Copy only the API key string from the snippet.

Example API key format

api_key="xxxxxxxxxxxxxxxxxxx"

---

🧩 Adding the API Key to the Code (Mandatory)

Before executing any training script, the API key must be added.

Example: H-CoAtNet training script

Open:

proposed_method/train_h_coatnet.py

Add the configuration block near the top of the file:

#  Configuration
API_KEY = "PASTE_YOUR_KEY_HERE"

Then initialize the dataset:

from roboflow import Roboflow

rf = Roboflow(api_key=API_KEY)

Important notes

• Use the same dataset version for all baseline and proposed models.

---

6. Training and Execution

Proposed Method (H-CoAtNet)

python proposed_method/train_h_coatnet.py

Baseline Models

python baselines/train_cnn.py
python baselines/train_efficientnet.py
python baselines/train_vit.py
python baselines/train_swin.py
python baselines/train_coatnet.py
python baselines/train_gft.py

All models use identical dataset splits, preprocessing, and evaluation protocols.

---

7. Experimental Protocol (Reproducibility)

• Optimizer: Adam
• Epochs: 30
• Dropout: 0.2
• Weight decay enabled
• No external pretraining (trained from scratch)
• Fixed random seeds

Hardware

• Apple MacBook Pro (M3 Pro, 18 GB RAM)
• Google Colab (verification only)

No TPU-specific optimizations are used.

---

8. Evaluation Metrics

• Accuracy
• Macro-averaged Precision, Recall, F1-score
• Weighted F1-score

Macro-averaged metrics are emphasized due to class imbalance inherent in rare disease datasets.

---

9. Results Summary

Model	Accuracy	Macro F1	Weighted F1
H-CoAtNet (Ours)	90.51%	0.8605	0.9024
Swin Transformer	82.91%	0.7477	0.8150
GFT	82.28%	0.7701	0.8221
CoAtNet	74.68%	0.6517	0.7463
Vision Transformer	72.15%	0.6310	0.7103
CNN	69.62%	0.6085	0.6889
EfficientNet-B0	66.46%	0.5938	0.6675

---

10. Ethical Considerations

• No patient-identifiable data is used
• Images are anonymized and sourced from publicly available materials
• Intended strictly as a clinical decision-support system, not a standalone diagnostic tool

---

11. Contact

Anandhu P. Shaji Email: reach.anandhu.me@gmail.com

---

⚖️ Legal Notice & Copyright

Copyright © 2026 The Authors. All Rights Reserved.

This repository, titled “Hierarchically Enhanced Hybrid Learning for Ichthyosis Classification (H-CoAtNet)”, and all associated materials — including but not limited to source code, experimental pipelines, benchmark datasets, execution logs, research documentation, and the accompanying manuscript — are provided solely for the purposes of peer review, validation, and reproducibility assessment in connection with the submitted work:

“Hierarchical Hybrid Learning: Enhanced Classification of Ichthyosis Variants in Dermatological Images Using H-CoAtNet”
Athul Joe Joseph Palliparambil, Anandhu P Shaji, Rajeev Rajan (Under Review, 2025)

This repository constitutes the official research codebase for a hierarchically enhanced hybrid convolution–transformer framework designed for multi-class Ichthyosis subtype classification from dermatological imagery.
This is NOT an open-source release.

Restrictions

 • The following actions are strictly prohibited without prior explicit written consent from the authors:
 • Reuse of any portion of the codebase in external projects, systems, or products
 • Redistribution of this repository or its contents, in full or in part
 • Modification, adaptation, translation, or creation of derivative works
 • Deployment of the framework or its components in production or clinical systems
 • Citation or reference to unpublished results prior to formal publication
 • Utilization of repository contents for training machine learning models

Legal Notice
Unauthorized use may constitute copyright infringement, intellectual property violation, and/or misappropriation of unpublished academic material.

Permitted use: Reviewers assigned by the programme committee may read, compile, and run the code solely for the purpose of evaluating the submitted manuscript. No other use is permitted.

---