Deep Convolutional Neural Network for Two-Dimensional Medical Image Segmentation Using a Learned Curriculum

Radiologists greatly benefit from medical image segmentation algorithms when determining the severity of diseases, conducting clinical evaluations, and choosing the best course of therapy. Deep learning is quickly becoming the go-to method for medical picture segmentation because to both the proliferation of big data and improvements in medical imaging methods. The medical segmentation structure encounters various obstacles, such as a lack of labelled data, complicated picture representation, low contrast variation, high structural similarity among infection as well as background, and irregularities related to texture and morphology within homogeneous infected areas. For the purpose of training convolutional neural networks to segment deformable paired 3D medical images, we investigate several curriculum learning approaches. We believe our approach to training medical image segmentation algorithms via curriculum learning, which involves beginning with a simple setup and progressively increasing its complexity, is the first of its kind to aim at improving performance. First, we take a look at two methods for learning curricula that are already out there: curriculum dropout along with curriculum through smoothing. But we provide a fresh and easy way to do curriculum: start with intentionally blurry photographs and work your way up to crisper ones throughout training. Curriculum learning can outperform traditional training methods, according to our studies using a state-of-the-art deep learning model. Among the evaluated curriculum learning methodologies, we also demonstrate that curriculum through input blur offers the highest accuracy against time trade-off.