Deep learning (DL) provides computational models of multiple processing layers to learn and represent data with multiple levels of abstraction. It is able to implicitly capture intricate structures of large-scale data and is ideally suited to some of the hardware architectures that are currently available. DL has recently achieved outstanding performance in academic and industrial fields, and become a vital utensil in a wide range of medical image computing tasks, including cancer detection, tumor classification, vessel segmentation, etc.
While DL models give impressively high predictive accuracy, they are recognized as black boxes with deep and complicated layers. In the meantime, DLs have been recently reported as defenseless to spoofing with elegant hand-designed input samples. This principally takes place in the medical image computing field, where a single incorrect prediction might be very detrimental, and the trust on the trained DL model and its capacity to deliver both efficient and robust data processing must be pledged. Therefore, understanding how the DL models work, and thus creating explainable DL models, have become an elemental problem.
Currently, it is still not clear what information must be delivered to DL models and how DL models work to warrant a rapid, safe, and robust prediction. Hence, experts/users request to know the latest research advances of explainable deep learning (XDL). This critical research topic will bring new challenges and opportunities to the AI community. The purpose of this special issue aims to provide a diverse, but complementary, set of contributions to demonstrate new developments and applications of explainable deep learning, to solve problems in medical image computing. The ultimate goal is to promote research and development of explainable deep learning for multimodal biomedical images by publishing high-quality research articles and reviews in this rapidly growing interdisciplinary field.
A total of more than 30 manuscripts were submitted and eight were selected based on a robust peer-reviewed process. The eight articles are authored by researchers from world-wide universities and reflect state of advances on explainable deep learning for medical image computing.
A novel local constraint and label embedding multi-layer dictionary learning model called LCLM-MDL is proposed in the study by Ni et al. to promote the discriminability of the classification model for sperm head classification.
Chen et al. propose a novel convolutional neural network (CNN)-based deep active context estimation (DACE) framework, which leverages the unlabelled neighbors to progressively learn more robust feature representations and generate a well-performed classifier for COVID-19 diagnosis.
Traditional deep learning models use Bilinear Interpolation when processing classification tasks of multi-size medical image data sets, which cause the loss of information of the image, and then affect the classification effect. In response to this problem, Liu et al. propose a solution for an adaptive size deep learning model.
The accurate diagnosis of COVID-19 is an emergent task as to stop the spreading of the virus. Lu et al. propose to leverage image feature fusion for the diagnosis of COVID-19 in lung window computed tomography (CT). The predictions of the three randomized neural networks (RNNs) are ensembled to get a more robust classification performance.
Alizadehsani et al. propose a Semi-supervised Classification using Limited Labelled Data (SCLLD) relying on Sobel edge detection and Generative Adversarial Networks (GAN) to automate the COVID-19 diagnosis. Their system is capable of learning from a mixture of limited labelled and unlabelled data, where supervised learners fail due to a lack of sufficient amount of labelled data.
In the study of Kumar et al., magnetic resonance (MR)-based pre-processed brain images are received by the Subtractive Spatial Lightweight Convolutional Neural Network (SSLW-CNN) model, which includes additional operators to reduce the complexity of classification.
Parkinson's disease is the second most common neurodegenerative disorder. As the cardinal manifestation, hypomimia, referred to impairments in normal facial expressions, stays covert. Su et al. propose a Semantic Feature based Hypomimia Recognition network, named SFHR-NET, to recognize hypomimia based on facial videos.
Finally, in the paper of Qi et al, an explainable graph feature CNN named WTRPNet is proposed for Epileptic electroencephalogram (EEG) classification. Experimental results show that their algorithm has achieved an accuracy of 99.67% in classification of focal and nonfocal epileptic EEG, which proves the effectiveness of the classification and detection of epileptic EEG.
As a final remark, we wish this Special Issue “Explainable Deep Learning for Medical Image Computing” can contribute to the field of explainable deep learning for medical signal computing, and it may benefit broader readers of researchers, practitioners, and students who are interested in related topics. We would like to thank the authors for their contributions to this Special Issue. We also thank the journal—ACM Transactions on Multimedia Computing, Communications, and Applications—for their supports for publications of this Special Issue.
Yu-Dong Zhang
School of Informatics, University of Leicester, UK
Juan Manuel Gorriz
University of Granada, Spain/Cambridge University, UK
Zhengchao Dong
Columbia University, USA
Guest Editors
