- 4 min read

The topic of Artificial Intelligence (AI) is on the tip of everyone’s tongue in healthcare, especially those involved in radiology. When one thinks of AI in radiology, initial thoughts may consist of new imaging algorithms or futuristic machines assisting doctors. However, one area that may not immediately be considered is that of publishing all the new research. Radiological journals must also consider this new territory, answering questions regarding how journals can appeal to authors on the topic of AI, the structure of these articles and how they fit into classic journals, and how Editors-in-Chief can encourage authors to submit papers on AI to their journals. We spoke with Professor Zhi-Cheng Li of the Research Center for Artificial Intelligence in Medicine at the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences to explore his thoughts on AI and its implications in the world of radiological journals.

Although AI may be a popular topic in the world of radiology, there is the issue of material and information that may be lost in translation. This may be due to the uncommon terminology between scientists and clinicians, which makes it more difficult for scientists to translate their study into a clinical perspective, and vice versa. “It is a big open question,” says Professor Zhi-Cheng Li. “For readers, one of the most important things is that they can benefit from reading an article.” Li includes two groups to whom this applies: First, MD or PhD candidates, or people who are not too familiar with AI techniques, require studying the know-how in designing an AI processing pipeline. However, it may not be that straightforward for them regarding access to this knowledge. The second group includes scientists or clinicians who are interested in papers of quality that provide useful knowledge, which provide insight and may inspire future studies.

Professor Li offers several possible solutions for journals regarding the issue of boosting the topic of AI among their readership: the relative journals can invite famous experts on a variety of areas within AI research to periodically write featured articles. This may significantly increase the impact. Alternatively, a special topic focusing on common problems in AI research can be created, which could be attractive and useful. “For example,” says Li, “we can focus on challenges of reproducible research in AI modelling, or designing some checklist or even score system to evaluate if an AI study is well designed.” Li explains further, “As another example, we can focus on the image normalization or data standardization issues, which are of essential importance. Other problems of common interest include: designing a radiomics pipeline, designing a convolutional neural network (CNN)-based learning framework, explanation of an AI model, survival prediction using AI model, AI-based radiogenomics analysis for tumour precision medicine, etc.”

Another solution could be that the journal publishes back-to-back papers, offering the readers the chance to view the issues from the perspectives of clinicians and scientists alike. For example, a topic like AI algorithms and chest imaging could be written from the point-of-view of the clinician and scientist, showing a direct compare and contrast, so the reader can more easily grasp the concepts in front of them.

In the AI community, holding challenges is an important way to increase the impact. Li explains that some societies have held many challenges in the area of AI-based medical image analysis; additionally, there have been challenges on brain tumour segmentation and survival prediction. Each challenge focuses on one specific application, where standardized medical imaging data and gold standard labels are provided and can be accessed freely. The aim of these challenges is to establish a benchmark for the evaluation of different AI models. Li suggests that journals could organize some similar type of challenge, which could be held annually or at selected congresses and meetings, with selected papers being published in a special issue of each respective journal.

Structuring articles on the topic AI

With the submission of articles covering areas within AI, such as deep learning and convolutions neural networks (CNN), the question of article structure has come up. Li is of the mindset that there is a clear need to invent a new template for papers relating to AI. “In my opinion, the background, the model, the statistical analysis, and the data constitute four major aspects of an AI paper,” says Li. “In the background, the authors should clearly state what the clinical requirement is, what type of problems or challenges the clinician faces, what the disadvantages of traditional models are, and what the hypothesis, basic idea, or objective is.” A consistent article structure may be helpful and increase submissions from authors, thus resulting in more articles published on the topic of AI.

In addition to the importance of article structure, Li discussed how essential it is to have an evaluation checklist for AI papers. “For example, for radiomics study, there is an evaluation checklist/online scoring system that helps authors conduct solid radiomics research (http://www.radiomics.world/). However, radiomics is based on traditional machine learning techniques. So far, it seems that there is no such checklist focusing on deep learning/CNN-based studies.” Therefore, the development of evaluation checklists to cover the various sub-topics of AI is of great importance.

The role of the Editor-in-Chief

Although there are many researchers working on AI-related papers, Li believes they are hesitant to submit their work to journals covering the field of radiology. This may be due to the fact that these journals are more traditional in nature and could refuse submissions on AI, as the editors and reviewers could consider these new submissions to be “too innovative.” Li explains that it may be more realistic for such authors to submit a “not-so-innovative” paper to increase their chances of having the paper accepted and published. Li also suggests Editors-in-Chief taking a more proactive role. “An Editor-in-Chief who invites and publishes several solid AI studies could provide a clear signal of encouragement for innovative papers,” says Li. Authors who see journals publishing more articles on the topic of AI would be more likely to submit their articles for review.

Although the idea of a journal dedicated solely to the topic of AI research has been discussed, Li thinks that this runs an unpredictable risk. “For authors, it would be better for existing, prominent journals to publish their AI papers,” as opposed to the work that goes into the creation of an entirely new journal.

 

Professor Zhi-Cheng Li works at the Research Center for Artificial Intelligence in Medicine, Institute of Biomedical and Health Engineering at the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences.

 

Latest posts