Bounding Box Based Weakly Supervised Deep Convolutional Neural Network for Medical Image Segmentation Using an Uncertainty Guided and Spatially Constrained Loss

Golnar K. Mahani, Ruizhe Li, Nikolaos Evangelou, Stamatios Sotiropolous, Paul S. Morgan, Andrew P. French, Xin Chen

Abstract: In this paper, we propose a weakly supervised deep convolutional neural network for medical image segmentation using an uncertainty guided and spatially constrained loss, which only requires bounding box annotations for model training. We utilise predictive uncertainty estimation during training to guide the model learning from the image region with high predictive confidence. Additionally, a conditional random field (CRF) based local spatial constraint is incorporated to the loss function, which regularises the predicted labels of a local region. This CRF loss term is independent to the training labels (bounding box annotation), which prevents the model over-fitted to the bounding box annotation. We evaluated our method on a public dermoscopic dataset containing different types of skin lesions. Our method achieved superior performance in comparison with the state-of-the-art learning based (DeepCut) and non-learning based (GrabCut) methods in terms of dice coefficient. The code is available on Github (https://github.com/golnarkmahani/Weakly-Supervised-Segmentation).