An AI system designed to identify diseased regions on cucumber leaves achieved a Dice score of 0.923 on a held-out image test set.
The authors caution that this does not establish reliable performance in other greenhouses.
What researchers studied
Researchers in Kazakhstan developed LARU-Net, a computer-vision architecture that separates visible lesions from the rest of a cucumber-leaf image.
It also estimates disease severity on a continuous scale from zero to four.
The system combines multiple image-processing components, including an attention-guided decoder and boundary-refinement module.
Its purpose is to identify where lesions appear, not simply label a whole leaf as healthy or diseased.
The researchers evaluated the model on images from a greenhouse collection.
Their reported held-out test partition contained 1,080 images.
The study also examined segmentation overlap, lesion-boundary accuracy and error in predicted severity scores.
What they found
On the held-out test partition, the reported Dice score was 0.9230 and intersection-over-union was 0.8660.
Both metrics compare predicted lesion regions with reference annotations, with higher scores indicating closer agreement.
The model also achieved a boundary F1 score of 0.9350 and precision of 0.9460.
For severity estimation, mean absolute error was 0.29 points, root mean squared error was 0.37 and reported R-squared was 0.90.
Importantly, the paper distinguishes these test-set results from an earlier highlighted Dice value of 0.9456, which it identifies as a validation score.
Mixing validation and held-out test performance can make model results appear stronger than they are.
What it means
Automated lesion mapping could eventually help growers monitor disease progression and decide which plants require closer inspection. A severity score may be more useful for tracking changes than a simple yes-or-no disease label.
The research is also an example of why evaluation details matter in agricultural AI.
High image-level scores are promising, but a model deployed on a new farm must cope with changing lighting, plant varieties, camera angles and disease appearances.
The reported results support further testing within this research setting; they are not evidence that the system is ready for autonomous crop-management decisions.
Important limitation
The authors report only single-run point estimates from one greenhouse image collection.
They did not retain complete plant and session identifiers, repeated-run logs or an independent external dataset in the study archive.
Images were split at image level, and independence between training and test images from the same plant, leaf or session could not be verified.
These gaps prevent confident claims about cross-greenhouse or cross-season generalisation.
Source Information
Study: LARU-net: lesion-aware residual attention U-net for binary cucumber leaf lesion segmentation and continuous severity estimation
Authors: Batyrkhan Omarov; Nurbibi Imanbayeva; Arman Ibrayeva
Journal / Source: Frontiers in Artificial Intelligence
Published: 5 November 2026
