Classify signs of retinopathy in image features
Use 19 extracted Messidor image descriptors to classify the presence of recorded diabetic-retinopathy signs. The Debrecen benchmark contains 1,151 rows and the saved project tests 230 of them.
1. Clinical question and intended use
The research question is whether image-level lesion and anatomical descriptors distinguish the provided signs-of-retinopathy label. This is an image-feature classification experiment, not a forecast of a future retinopathy event.
Image descriptors
Use the feature extraction defined by the dataset.
Operating point
Inspect confusion counts at threshold 0.5.
Evidence boundary
Separate internal classification from clinical utility.
2. Cohort, measurements and endpoint
The UCI Diabetic Retinopathy Debrecen data include image quality and prescreening flags, microaneurysm and exudate descriptors, normalized anatomical measurements and an AM/FM classification result. Class 1 represents signs of retinopathy; class 0 represents their absence in the source annotation.
Source: Diabetic Retinopathy Debrecen. Dataset license: CC-BY-4.0. The downloadable ZIP includes the adapted data and attribution notices.
| Dataset measure | Saved value |
|---|---|
| Analysis unit | image-feature record |
| Records | 1,151 |
| Raw variables | 20 |
| Encoded model inputs | 19 |
| Model outputs | 1 |
| Training roles | 691 |
| Validation / selection roles | 230 |
| Testing roles | 230 |
| Unused roles | 0 |
| Field | Role | Type | Categories |
|---|---|---|---|
| quality | Input | Binary | 0, 1 |
| pre_screening | Input | Binary | 0, 1 |
| ma1 | Input | Numeric | |
| ma2 | Input | Numeric | |
| ma3 | Input | Numeric | |
| ma4 | Input | Numeric | |
| ma5 | Input | Numeric | |
| ma6 | Input | Numeric | |
| exudate1 | Input | Numeric | |
| exudate2 | Input | Numeric | |
| exudate3 | Input | Numeric | |
| exudate4 | Input | Numeric | |
| exudate5 | Input | Numeric | |
| exudate6 | Input | Numeric | |
| exudate7 | Input | Numeric | |
| exudate8 | Input | Numeric | |
| macula_optic_disc_distance | Input | Numeric | |
| optic_disc_diameter | Input | Numeric | |
| am_fm_classification | Input | Binary | 0, 1 |
| class | Target | Binary | 0, 1 |
3. Model
The model has 19 encoded inputs and 1 outputs. No architecture-selection experiment is recorded in this project. The diagram shows the topology used by the saved model.
| Layer | Input shape | Output shape | Activation |
|---|---|---|---|
| Scaling | 19 | 19 | |
| Dense | 19 | 3 | Tanh |
| Dense | 3 | 1 | Sigmoid |
Output values are uncalibrated model scores. For binary evaluation, the saved positive class is 1.
4. Training strategy
The saved training configuration uses QuasiNewton with WeightedSquaredError.
Quasi-Newton method results
| Measure | Value |
|---|---|
| Epochs number | 190 |
| Elapsed time | 00:00:00 |
| Stopping criterion | Maximum validation error increases |
| Training error | 0.605 |
| Validation error | 0.691 |
5. Model selection and baseline
No model selection experiment is recorded for this version. The validation subset guides fitting where a training report is present; it is distinct from the held-out test rows.
On this test subset a majority-class baseline would correctly classify 51.74% of the records. This baseline does not detect both classes and is not an optimized model.
6. Clinical validation
The figures and tables below refer to the current project’s saved testing analysis. The subset uses testing role 2; it contains 230 source records.
ROC AUC describes ranking on this testing subset. Any optimal threshold shown in the saved ROC report was selected descriptively on that same subset; it is not an independently validated operating policy.
These operating-point metrics are calculated from the saved confusion counts at threshold 0.5. The positive-label coding is stated in the model section.
Confusion table
| Measure | Predicted positive | Predicted negative | Total |
|---|---|---|---|
| Actual positive | 84 (36.5%) | 35 (15.2%) | 119 (51.7%) |
| Actual negative | 14 (6.1%) | 97 (42.2%) | 111 (48.3%) |
| Total | 98 (42.6%) | 132 (57.4%) | 230 (100.0%) |
| Test measure | Value |
|---|---|
| Testing records | 230 |
| Positive cases | 119 |
| Positive prevalence | 51.74% |
| Accuracy | 78.70% |
| Sensitivity / recall | 70.59% |
| Specificity | 87.39% |
| Precision / PPV | 85.71% |
| F1 score | 0.774 |
Area under curve
| Measure | Value |
|---|---|
| Area under curve | 0.854 |
7. Workflow and reproducibility
Open the downloaded project in Neural Designer, inspect the dataset roles and preprocessing, then review the saved task report. Use the same input schema and category order when calculating outputs. The ZIP contains the exact current .nd, its source data and the applicable dataset notices.
Research workflow: eligible record or image → measurement and schema checks → model score → expert review → confirmatory clinical method where appropriate. No model output should be used as an autonomous diagnosis or treatment instruction.
8. Safety, generalizability and governance
The supplied CSV does not provide a patient identifier for auditing patient-separated splits. Results therefore describe the saved internal record split, not independent patient-level generalization. No external validation or calibration has been established; ophthalmology specialist review is required.
References
- Diabetic Retinopathy Debrecen. Balint Antal; Andras Hajdu. 10.24432/C5XP4P
- Dataset terms: Creative Commons Attribution 4.0 International. Full attribution and transformations are included in
LICENSES/DATASET-LICENSE.txt. - Current Neural Designer project and saved task report, snapshot 6 October 2026.
