On the fragility of neural architecture search: the role of overfitting and task complexity in medical image analysis.

The fMRI literature has accumulated many studies that investigated the neural correlates of taste and food stimuli. As a recent meta-evaluation pointed out, tens of meta-analyses have been published to include original studies that used numerous forms of food stimuli, ranging from food images, tastants (taste solutions), food odors, real food items, to mental imagery (). Meta-analyses showed that food images induced the most extensive and robust activation across the brain especially the visual cortex (), whereas taste and smell stimuli induced stronger activation in the insula, the primary taste cortex (). It demonstrated the complexity and multi-faceted nature of the cerebral processing of food. In terms of tastants, the fMRI literature has been relatively standardized in the selection of chemicals to provide the five basic tastes, with popular examples being sucrose for sweetness, citric acid for sourness, quinine for bitterness and sodium chloride for saltiness (; ). However, the situation for food images seemed to be different. The descriptions of food images seemed to varied, with one study describing the stimuli as “roasted chicken, hamburger, chocolate cake, strawberries, etc.” to another study with a simple description as “food” [see () for the examples]. Since multiple food images may be used in a single study, authors may not list out everything in the article for simplicity or due to page restrictions. Hence, a viable alternative may be to disclose the sources of food images. One popular picture database is the International Affective Picture System (IAPS), developed by the National Institute of Mental Health (NIMH) Center for Emotion and Attention at University of Florida. It seems that the release of IAPS was officially introduced in some technical documents instead of a journal publication (). IAPS includes over 1,000 color photos that depict a wide variety of events and objects (including food images), and were rated by large standardized groups for evoked feelings of arousal, dominance, and pleasure (). Notwithstanding, IAPS was not specifically designed for food studies, such as the food items presented may have limitations. Food image datasets have since been developed, such as Food-pics (). It has a broader coverage of food items, such as vegetables versus meat versus fruits, high versus low-calorie food, sweet versus salty food, and whole versus processed food. Differences in visual stimulus sets may influence neural activation patterns, particularly in regions such as the visual cortex and reward-related circuits, making transparent reporting of stimulus materials critical for interpreting and comparing fMRI findings (; ). It was largely unclear if food studies have used food images from established datasets and exactly from which datasets. This could be particularly relevant to fMRI studies for which small methodological differences may lead to significantly different results (; ). Given this context, the purpose of this meta-research study is to characterize reporting practices and sourcing of food images in fMRI studies, and to evaluate their implications for reproducibility and comparability of neuroimaging findings.