Effimix

Problem

Classifying gastrointestinal diseases from endoscopy images is hard: the differences between conditions are subtle and labelled data is limited, so a single backbone tends to either overfit or miss disease-specific structure.

What I built

Effimix, a CNN that fuses features from a pretrained EfficientNet-B0 with a purpose-built branch using squeeze-and-excitation blocks and self-normalising activations. It was evaluated on the HyperKvasir endoscopy dataset.

Key decision or trade-off

Feature fusion rather than fine-tuning one network. Combining a pretrained backbone with a custom branch lets the model keep general visual features while still learning what is specific to endoscopy images.

Result

The method surpassed prior benchmarks on the task and was published in Diagnostics (MDPI), 2022, volume 12, article 2316.