Search

The latest deep neural networks for medical segmentation typically utilize transposed convolutional filters and atrous convolutional filters for spatial restoration and larger receptive fields, leading to dilution and inconsistency of visual semantics. To address such issues, we propose a novel attentional up-concatenation structure to build an auxiliary path for direct access to multi-level features. In addition, we employ a new structural loss to bring better morphological awareness and reduce the segmentation flaws caused by the semantic inconsistencies. Thorough experiments on the challenging optic cup/disc segmentation, cellular segmentation and lung segmentation tasks were performed to evaluate the proposed methods. Further ablation analysis demonstrated the effectiveness of the different components of the model and illustrated its efficiency. The proposed methods achieved the best performance and speed compared to the state-of-the-art models in three tasks on seven public datasets, including DRISHTI-GS, RIM-r3, REFUGE, MESSIDOR, TNBC, GlaS and LUNA. • A deep learning model, Up-Net, is proposed to overcome semantic inconsistency issue. • A lightweight up-concatenation structure and metric learning strategy is proposed. • We validated Up-Net in 3 challenging medical segmentation tasks across 7 datasets. [ABSTRACT FROM AUTHOR]

Transfer learning from natural image datasets, such as ImageNet, is common for applying deep learning to medical imaging. However, the modalities of natural and medical images differ considerably, and the reason for the latest medical research preferring ImageNet to medical data is questionable. In this study, we investigated the properties of medical pre-training and its transfer effectiveness on various medical tasks. Through an intuitive convolution-based analysis, we determined the modality characteristics of images. Surprisingly, medical pre-training showed exceptional performance for a classification task but not for a segmentation task since medical data are visually homogeneous and lack morphological information. Using data with diverse modalities helped overcome such drawbacks, resulting in medical pre-training achieving performance comparable to pre-training with ImageNet with considerably fewer samples than ImageNet for both aforementioned tasks. Finally, a study of learned representations and realistic scenarios indicated that while ImageNet is the best choice for medical imaging, medical pre-training has significant potential. • We study the transfer learning effectiveness based on medical and natural images. • We evaluate medical pre-training on downstream medical image analysis tasks. • We reveal the shortcomings and future potential of medical pre-training. [ABSTRACT FROM AUTHOR]