Gastrointestinal Polyp Segmentation (Master's project)

What are gastric lesions and why segmentation is needed?

• Gastric lesions are abnormalities or damage found in the stomach lining or the tissues of the stomach. These lesions can be a result of various conditions or diseases and may vary in appearance and severity. Several types of gastric lesions include polyps, ulcers, neoplasms etc. This project focus on gastric polyps which are usually found incidentally during an upper endoscopy. While they are generally benign, they can potentially develop into cancer.
• Research in this field has the potential to help reduce the polyp miss rate and thus improve examination quality.

Data

Dataset: Kvasir-SEG dataset
Source: https://datasets.simula.no/kvasir-seg/

It contains 1000 polyp images and their corresponding ground truth.

Architecture

I Used a simple U-net, with 3-channel input and 1-channel output. Check notebook for the model arch.
Architecture based on paper: https://arxiv.org/pdf/1505.04597.pdf

Training

Trained for 120 epochs. Check Notebook for the whole training in PyTorch.
Criterion - DiceLoss
Optimizer - Adam

Results

Loss on test Data -
Dice Loss: 1.0509
IoU Loss: 1.0000
BCE Loss: 0.0509

Predictions-

Code

The notebook folder contains the data preperation, architecture and training pipeline code.