Segmantation of polyps in colonoscopic images

Polyps are abnormal tissue outgrowths that, although initially benign, can evolve into malignant lesions over time. Colorectal cancer arises upon the malignant transformation of such polyps; thus, early detection and appropriate treatment are essential for prevention. Clinical assessment of colorectal polyps by expert endoscopists is susceptible to delays, high costs and human error. Computer-aided diagnostic systems address these limitations by automating the detection process, thereby reducing both time expenditure and false-negative or false-positive rates. A comprehensive literature review was conducted to examine current methods for automatic colorectal polyp segmentation. Subsequently, the DeepLabv3+ architecture—renowned for its robust and precise segmentation capabilities—was evaluated using four convolutional backbones: DenseNet, ResNet50, SqueezeNet and VGG16. Each DeepLabv3+ variant was tested on the Kvasir-SEG and CVC-ClinicDB datasets, employing seven performance metrics: mean Dice coefficient, mean Intersection over Union, accuracy, recall, specificity, precision and mean squared error. On the Kvasir-SEG dataset, the best overall performance was achieved using the DenseNet backbone. Specifically, the model obtained a mean Dice score of 0.858, a mean IoU of 0.850, accuracy of 0.948, recall of 0.824, precision of 0.896, and a mean squared error of 0.045. In terms of specificity, the highest result (0.984) was observed with the VGG16 backbone. Another component of this thesis involved evaluating segmentation performance using the U-Net architecture with different loss functions. Four models were trained using Binary Cross-Entropy, Tversky, Dice, and Composite loss functions. The performance of these models was thoroughly evaluated on the CVC-ColonDB and Kvasir-SEG datasets. Experimental results demonstrated that the Binary Cross-Entropy loss function consistently yielded the highest segmentation accuracy across both datasets. The highest accuracy was 0.9761 on the CVC-ColonDB dataset and 0.9313 on the Kvasir-SEG dataset. These findings highlight the critical role of loss function optimization in achieving high segmentation accuracy and model stability.