Release Subtitle:
Scientists use a deep learning approach to look at internal diseases in persimmon fruits, but with a different perspective

Release Summary Text:
Scientists from Japan have used a ‘feature visualization’ based explanatory deep learning approach to diagnose symptoms of “hetasuki” disorder in persimmon fruit, that have long eluded even professionals. In their latest paper, based on the relevant features detected by the artificial intelligence, they recommend uneven coloring patterns as a possible indication of disorder. These findings showcase a novel use for deep neural networks in plant physiology.

Full text of release:
While originally modeled on the human brain, deep neural networks (DNNs) today can surpass our capabilities in many activities; a DNN model named ResNet50, for instance, has been shown to classify images better than a human eye. Consequently, DNNs have been deployed for activities that would otherwise require a trained human eyesight, such as identifying symptoms of diseases in plants or telling moss species apart. But what about identifying disorders that present no apparent symptoms (such as an internal damage)? Till now, DNNs have found little use in such cases.

A possible reason might be that until recently, DNNs have largely been a “black box” that only makes predictions but doesn’t tell you how they are made. This meant having no information on what regions of the image contributed to the prediction; one couldn’t tell if DNNs would provide any insight that couldn’t be discerned by a trained professional. Fortunately, various feature visualization methods are now available that can significantly alleviate the “black box” situation, telling us if DNNs can see things a “professional eye” cannot.

In a new study published in Plant & Cell Physiology, scientists from Japan decided to employ DNNs for diagnosing an internal disorder in persimmon fruits called “hetasuki” ( Japanese for “calyx-end cracking”) signified by appearance of cracks around the calyx (leafy part on top of the fruit). “Hetasuki has been well classified into five grades; the higher degrees not only visually spoil the fruit quality, but often become a trigger of quick softening. However, no clear outer symptoms or biomarkers have yet been identified. Moreover, only very few experts with decades of experience in persimmon cultivation can identify hetasuki from its external appearance. We, therefore, wanted to examine the utility of DNN for diagnosis of internal reactions in fruit,” explains Associate Professor Takashi Akagi from Okayama University, who led the study.

The scientists collected a total of 3,173 colored images of the fruit apex and fed them into pre-trained DNNs for a binary (positive or negative) classification between different pairs of cracking levels. They chose five different DNN models for the classification and compared their performance against one another. To visualize the relevant image regions for the diagnosis, they applied feature visualization methods on the relatively simpler VGG16 model and compared their outcomes.

The team found that all the five DNN models performed well in classifying the fruits for the level pairs considered except for the pair 0-3 vs. 4, which they attributed to a severe imbalance of samples (level 0-3 occupied >98% of the samples). Among the models, InceptionResNetV2 showed the highest accuracy (>90%); however, due to its highly complicated layer structure, scientists chose VGG16 for feature visualization. Nevertheless, the relevant regions were often inconsistent among the applied methods. However, looking at the distribution of relevance values as a function of distance from the outer contour of fruit, the relevance regions were located either around the apex or close to the periphery of the fruit.

Interestingly, while the relevance of periphery can be explained based on uneven coloring observed in cracked fruits, that of the apex has no such visual guide and is therefore a feature picked up only by an AI! With this finding, scientists are looking forward to more extensive applications of AI in the field of plant physiology. “Our study shows that DNN can act as an ‘artificial professional eye’ for fruit physiology diagnosis. In the future, maybe in just 10 years, when most of the old experts on persimmon fruit cultivation will be retired, this would help the situation greatly. Instead of a competition between humans and AI, I look forward to an ‘interaction’ between them that would help overcome seemingly impossible challenges,” concludes Dr. Akagi.

Maybe having AI outperform humans isn’t such a bad thing when it can serve to push the limits of our abilities further!

Reference:
Title of original paper: Explainable Deep Learning Reproduces a ‘Professional Eye’ on the Diagnosis of Internal Disorders in Persimmon Fruit
Journal: Plant & Cell Physiology
DOI: 10.1093/pcp/pcaa111

Contact Person: AKAGI Takashi
Dr Akagi is Associate Professor at the Graduate School of Environmental and Life Science at Okayama University, Japan. He conducts research on diverse control mechanisms for sex determination and expression in floral organs of horticultural crops, as well as on the shape and physiological disorders of fruits. He also strives to clarify the factors that contribute to the quality judgment and preference of fruits and vegetables assisted by deep learning (AI) technology. Dr. Akagi has 47 publications to his credit, with over 1000 citations.

Contact: E-mail: takashia(a)okayama-u.ac.jp
For inquiries, please contact us by replacing (a) with the @ mark.