01 / The question
Can an image reveal a disease stage?
The APTOS 2019 project classified diabetic-retinopathy severity on a five-stage scale. The task is ordinal: confusing neighbouring stages and confusing distant stages are not equivalent errors.
02 / The intuition
A mechanism worth testing.
Use a pretrained visual representation, adapt it to the retinal images and inspect what the model relies on. A saliency map is a question to investigate, not clinical validation.
03 / The work
Inside the method.
Explore each part of the approach.
01Prepare the images
The recorded pipeline included image preprocessing and augmentation to handle variation in retinal photographs.
02Adapt a pretrained CNN
Transfer learning was used for severity classification. The precise backbone and training configuration need the original experiment files.
03Inspect decisions
Grad-CAM maps were used to explore model attention. Quadratic weighted kappa was recorded as an evaluation criterion for the ordinal labels.
04 / The observations
What emerged.
The notes document transfer learning, augmentation and Grad-CAM. They mention a strong validation result, but the underlying predictions and competition record have not been supplied here, so no ranking or score is claimed.
This account is based on recorded project notes. Original reports, figures and datasets are not embedded here.
05 / The limits
Where the evidence stops.
Patient-level separation, class balance, preprocessing and external evaluation must be checked before interpreting performance. Attention maps alone do not establish that a model has learned clinically valid features.
The academic foundations