Meta Ordinal Regression Forest for Medical Image Classification With Ordinal Labels

The performance of medical image classification has been enhanced by deep convolutional neural networks (CNNs), which are typically trained with cross-entropy (CE) loss. However, when the label presents an intrinsic ordinal property in nature, e.g., development from benign to malignant tumor, CE loss cannot take into account such information allow for better generalization. To improve model generalization information, we propose a novel meta regression forest (MORF) method labels, learns relationship through combination network and differential meta-learning framework. merits proposed MORF come following two components: A tree-wise weighting net (TWW-Net) grouped feature selection (GFS) module. First, TWW-Net assigns each tree specific weight that is mapped corresponding tree. Hence, all trees possess varying weights, helpful alleviating prediction variance. Second, GFS module enables dynamic rather than fixed one was previously used, allowing random perturbation. During training, alternatively optimize parameters CNN backbone framework calculating Hessian matrix. Experimental results on datasets i.e., LIDC-IDRI Breast Ultrasound datasets, demonstrate superior performances our over existing state-of-the-art methods.