Many applications for the automated diagnosis of plant disease have been developed based on the success of deep learning techniques. However, these applications often suffer from overfitting, and the diagnostic performance is drastically decreased when used on test data sets from new environments. In this article, we propose LeafGAN, a novel image-to-image translation system with own attention mechanism. LeafGAN generates a wide variety of diseased images via transformation from healthy images, as a data augmentation tool for improving the performance of plant disease diagnosis. Due to its own attention mechanism, our model can transform only relevant areas from images with a variety of backgrounds, thus enriching the versatility of the training images. Experiments with five-class cucumber disease classification show that data augmentation with vanilla CycleGAN cannot help to improve the generalization, i.e., disease diagnostic performance increased by only 0.7% from the baseline. In contrast, LeafGAN boosted the diagnostic performance by 7.4%. We also visually confirmed that the generated images by our LeafGAN were much better quality and more convincing than those generated by vanilla CycleGAN. The code is available publicly at https://github.com/IyatomiLab/LeafGAN.

The aims of this study were to provide a quantitative assessment of the tumour area extracted by dermatologists and to evaluate computer-based methods from dermoscopy images for refining a computer-based melanoma diagnostic system. Dermoscopic images of 188 Clark naevi, 56 Reed naevi and 75 melanomas were examined. Five dermatologists manually drew the border of each lesion with a tablet computer. The inter-observer variability was evaluated and the standard tumour area (STA) for each dermoscopy image was defined. Manual extractions by 10 non-medical individuals and by two computer-based methods were evaluated with STA-based assessment criteria: precision and recall. Our new computer-based method introduced the region-growing approach in order to yield results close to those obtained by dermatologists. The effectiveness of our extraction method with regard to diagnostic accuracy was evaluated. Two linear classifiers were built using the results of conventional and new computer-based tumour area extraction methods. The final diagnostic accuracy was evaluated by drawing the receiver operating curve (ROC) of each classifier, and the area under each ROC was evaluated. The standard deviations of the tumour area extracted by five dermatologists and 10 non-medical individuals were 8.9% and 10.7%, respectively. After assessment of the extraction results by dermatologists, the STA was defined as the area that was selected by more than two dermatologists. Dermatologists selected the melanoma area with statistically smaller divergence than that of Clark naevus or Reed naevus (P=0.05). By contrast, non-medical individuals did not show this difference. Our new computer-based extraction algorithm showed superior performance (precision, 94.1%; recall, 95.3%) to the conventional thresholding method (precision, 99.5%; recall, 87.6%). These results indicate that our new algorithm extracted a tumour area close to that obtained by dermatologists and, in particular, the border part of the tumour was adequately extracted. With this refinement, the area under the ROC increased from 0.795 to 0.875 and the diagnostic accuracy showed an increase of approximately 20% in specificity when the sensitivity was 80%. It can be concluded that our computer-based tumour extraction algorithm extracted almost the same area as that obtained by dermatologists and provided improved computer-based diagnostic accuracy.