Related Experiment Videos
Optimised Deep Learning for Gastrointestinal Polyp Classification: A Controlled Benchmark of Five CNN and Transformer
Zhengsui Gu1,2, Hoda Anwar Ibrahim2, Wamadeva Balachandran2
1School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China.
Diagnostics (Basel, Switzerland)
|July 28, 2026
Summary
EfficientNet-B3 achieved the highest accuracy in classifying gastrointestinal polyps using deep learning. This study compared five models, finding EfficientNet-B3 superior for accurate polyp subtype classification in endoscopy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Gastroenterology
Background:
- Colorectal cancer (CRC) is a leading cause of cancer mortality globally.
- Colonoscopy polyp miss rates can reach 26%, and classification accuracy is operator-dependent.
- Accurate polyp subtype classification is crucial for determining appropriate treatment (resection vs. surveillance).
Purpose of the Study:
- To systematically compare five deep learning architectures for five-class gastrointestinal polyp classification.
- To provide clinically interpretable diagnostic insights using Grad-CAM visualization.
Main Methods:
- Five deep learning models (ResNet50, VGG16, EfficientNet-B3, DenseNet121, Vision Transformer (ViT-B/16)) were evaluated on the Kvasir Dataset V2 (5000 images, 5 classes).
- Models utilized ImageNet transfer learning with a redesigned classification head.
- Stratified five-fold cross-validation was employed, with Grad-CAM used for attention mapping.
Main Results:
- EfficientNet-B3 achieved the highest cross-validated accuracy (95.90 ± 0.35%), outperforming ViT-B/16 (95.12 ± 0.72%) and CNNs.
- All models demonstrated stable performance across cross-validation folds.
- A consistent misclassification pattern was observed between dyed-lifted polyps and dyed-resection margins, indicating visual ambiguity.
Conclusions:
- EfficientNet-B3 demonstrated superior accuracy for five-class gastrointestinal polyp classification under cross-validation.
- Vision Transformer (ViT-B/16) was a highly competitive second, outperforming other CNNs.
- Identified class confusion suggests potential need for higher-resolution imaging or expert re-annotation, not solely architectural improvements.