Kratos7270/ConvNeXt_Tiny_ordinal

ConvNeXt Tiny Ordinal for Diabetic Retinopathy Grading

Model Overview

This model is a ConvNeXt-Tiny based deep learning model fine-tuned for diabetic retinopathy (DR) grading on retinal fundus images.

The task is to predict diabetic retinopathy severity on an ordered 5-class scale:

• 0: No DR
• 1: Mild
• 2: Moderate
• 3: Severe
• 4: Proliferative DR

Unlike standard 5-class classification, this model uses an ordinal formulation, which is more suitable because DR grades have a natural severity order.

Architecture

• Backbone: ConvNeXt-Tiny
• Task type: Ordinal classification
• Output format: 4 ordinal logits for 5 ordered classes
• Prediction decoding: cumulative thresholding across ordinal outputs

The ordinal setup is designed to penalize large severity mistakes more appropriately than plain flat classification.

Intended Use

This model is intended for:

• academic experimentation
• coursework/demo projects
• research prototyping for diabetic retinopathy grading
• comparison with other retinal image models in a multi-model web interface

This model is not intended for direct clinical deployment or medical decision-making.

Training Data

The model was trained on the Kaggle Diabetic Retinopathy Detection dataset, which contains retinal fundus images labeled on a 0 to 4 severity scale.

Preprocessing

Typical preprocessing for this model:

• RGB retinal fundus image input
• resized to 224 × 224
• normalized using ImageNet-style mean and standard deviation

If custom preprocessing was used in the final experiment, update this section with the exact details.

Label Mapping

The model predicts one of the following DR grades:

Class	Severity
0	No DR
1	Mild
2	Moderate
3	Severe
4	Proliferative DR

Output / Decoding

The model produces 4 ordinal logits.
These are converted into probabilities using sigmoid, and final grades are obtained by applying thresholds and counting passed severity boundaries.

If custom thresholds are used, they should be stored alongside the model config or inference code.

Metrics

Recommended evaluation metrics for this task include:

• Quadratic Weighted Kappa (QWK)
• Accuracy
• Weighted F1-score
• Confusion Matrix

QWK is especially important because diabetic retinopathy grading is an ordinal prediction problem.

Limitations

• Performance depends heavily on image quality and preprocessing consistency.
• The model may not generalize well to data from different cameras, hospitals, or acquisition conditions.
• The model is for research/demo use and should not be used alone for diagnosis.

Ethical and Safety Notice

This model is not a medical device.
Predictions should not be used as a substitute for professional ophthalmology assessment.

Example Inference

Typical inference flow:

1. load the retinal image
2. resize and normalize it
3. run the model
4. apply sigmoid to ordinal logits
5. decode grade using thresholds
6. map predicted class to DR severity label

Files

This repo may include:

• model weights
• config.json
• preprocessing configuration
• label mapping
• threshold metadata
• inference helper code

Authors / Project Context

Developed as part of a diabetic retinopathy grading project for Applied Machine Learning coursework / research experimentation.