CholeNet
Undergraduate thesis — Gallbladder cancer classification using quad-fold parallel transfer learning, multi-model feature fusion, and LSTM. Achieved 99.37% accuracy.
CholeNet is a recurrent-convolutional hybrid ensemble framework developed as my undergraduate thesis at KUET. It addresses the challenge of gallbladder cancer classification from ultrasound images using a combination of deep learning techniques.
Key Contributions
- Quad-fold parallel transfer learning: Four pre-trained CNN backbones (VGG, ResNet, InceptionV3, DenseNet) run in parallel to extract complementary feature representations.
- Multi-model feature fusion: Features from all four branches are concatenated and fused to form a richer, more discriminative representation.
- LSTM-based sequential processing: Fused features are passed through LSTM layers to exploit sequential structure and temporal dependencies.
- 99.37% classification accuracy on the gallbladder cancer dataset — outperforming prior ensemble and single-model baselines.
Significance
Gallbladder cancer has poor prognosis due to late diagnosis. This framework provides a fast, non-invasive, automated diagnostic aid that can assist radiologists, particularly in resource-limited settings.
Status
Published in part at ICEEICT 2024 (IEEE). Extended version (CholeNet) is in preparation for Biomedical Signal Processing and Control (Elsevier).