Semantic Segmentation of Endoscopic Images using Transformers – Towards Trading off Dice and Precision

Detecting polyps in endoscopic images is very important for early detection of colorectal cancer, hence the importance of proper training of practitioners. Preparing a good set of images for training them is a challenging task, due to lack of comprehensive image sets. To address this, we explored a method based on Deep Learning to automatically detect polyps in colonoscopy images and generate the corresponding masks, such a solution providing the ability of turning unlabeled data into an important learning resource. We explored different augmentations techniques and used a FCBFormer based model, which was trained using different hyperparameters, managing to obtain two main versions that are noteworthy, one that is using the AdaBelief optimizer and the weighted sum between the Binary Cross-Entropy loss and the Dice loss as the loss function, which leads to a value of 0.93 for Dice and 0.99 for Precision, outperforming the baseline FCBFormer architecture. On the other hand, employing the Adam optimizer and Binary CrossEntropy loss leads to better Recall and Precision values (0.95 and 0.998 respectively) at the expense of a slight reduction of the Dice score from 0.93 to 0.92.