One of the main causes of death due to cancer worldwide is colorectal cancer . The most common method for locating, identifying, and removing colorectal polyps is a colonoscopy. When doing a colonoscopy, doctors frequently miss colorectal polyps because of their variety in size, shape, and similarity to surrounding tissue. Automated polyp segmentation can reduce missed rates and allow for early detection and treatment of colon cancer. Physicians may find it easier to identify polyps during a colonoscopy examination with the aid of Artificial Intelligence-based Computer Aided Diagnostics. In this work,we try to study polyp segmentation and build an efficient and accurate segmentation algorithm. We Use the U-net model along with some other pre-trained model as encoder of U-net model. We have achieved IOU score of 77.29% using mobilenetV2 as encoder for UNET model.
- Developed and tested U-Net and its variants (with encoders: MobileNetV2, ResNet50, DenseNet121, VGG19) for polyp segmentation using the Kvasir-SEG dataset.
- Applied data preprocessing (resizing, normalization, dataset split) and trained models with optimized settings (Adam optimizer, 400 epochs, learning rate scheduling).
- Achieved the best performance with U-Net + MobileNetV2, reaching a 77.29% IoU on test data.
- Compared models, analyzed generalization issues, and suggested improvements through data augmentation and hyperparameter tuning.
-
Provides an AI-assisted tool to help doctors detect polyps more accurately during colonoscopy, reducing missed cases of colorectal cancer.
-
Offers a lightweight, efficient model (MobileNetV2-U-Net) that reduces training time and computational cost.
-
Establishes a benchmark comparison of U-Net variants for medical image segmentation, guiding future research.
-
Contributes to early cancer detection and treatment, potentially improving patient survival rates.
Thanks For The Read