Medical image classification
Hybrid retinal-image classification
A manuscript study combining EfficientNetB0 and a Swin Transformer for diabetic-retinopathy severity classification on APTOS 2019.
The project studies whether local convolutional features and transformer context complement one another when grading retinal images.
Can a CNN–transformer hybrid improve severity classification while retaining interpretable image regions?
My contribution
- Co-authored the manuscript and contributed to the hybrid classification study.
- Worked with APTOS 2019, augmentation, focal loss and Grad-CAM interpretation.
Signature view
Trace the evidence.
Move across the research chain. The case study below remains complete and readable without this interaction.
Question
Combine local and global visual evidence
Retinal severity cues can be small and local while disease patterns span the fundus.
Method, as a sequence of decisions
- 01
EfficientNetB0 extracts local visual features.
- 02
A Swin Transformer models longer-range image relationships.
- 03
Focal loss addresses class imbalance and Grad-CAM inspects salient regions.
What the record actually supports
Dataset described
The manuscript’s retinal-fundus benchmark.
Reported average F1 improvement
Interpretation as relative vs. percentage-point improvement still needs confirmation.
Evidence: manuscript summary in the achievement record. The underlying paper was not supplied in the current workspace.
Choices and trade-offs
Why a hybrid?
Convolutions encode local texture efficiently, while windowed attention can model broader spatial structure.
What has not been established
- Publication status is unverified.
- The exact interpretation of the reported 9% improvement needs to be checked before formal citation.
- Dataset performance alone does not establish clinical generalisation.
What I would investigate next
- Q1How does the model perform across cameras, sites and demographic groups?
- Q2Are Grad-CAM regions stable under image perturbation?
No public paper or repository is attached yet. The case study is intentionally limited to the evidence in the supplied record.
Peer notes / DR hybrid
See a missed assumption?
Research suggestions, methodological critique and relevant collaboration ideas are particularly useful.