High-precision semantic segmentation of apple internal diseases based on CT imaging and deep learning

Moldy core and browning in apples are difficult to identify directly due to their hidden nature, making them easily overlooked during storage and sales, leading to reduced quality and economic losses. This study investigates the apple varieties ‘Ruixianghong’ and ‘Fuji,’ which are prone to disorders, as research subjects. Leveraging the non-destructive and high-penetration advantages of computerized tomography(CT) technology, the study focuses on two typical internal apple diseases—mold heart disease and browning—and proposes a method for segmenting and identifying internal apple diseases based on deep learning techniques. First, CT slice image data of apples are acquired through a CT imaging system and preprocessed. Two generative models—Denoising Diffusion Probabilistic Models and CycleGAN—are used to enhance the CT image data to address the issue of insufficient disease sample data. The proposed TransUNet-KDA model is based on the TransUNet architecture, enhanced with the CSWin Transformer module and a lightweight channel-space attention module. Ablation experiments demonstrate that the improved modules effectively enhance the accuracy and robustness of internal disease region segmentation in apples. In comparison experiments with a traditional thresholding approach and four classic semantic segmentation networks—UNet, FCN, Attention-UNet, and ResUNet—the improved model's segmentation performance significantly outperforms other models, achieving mIoU of 92.48%, OA of 91.78%, and mDice of 94.92% on the test set. Therefore, the semantic segmentation method proposed in this study based on the TransUNet-KDA model can effectively and accurately segment internal diseases in apple CT slices, providing new methods and technical support for rapid, non-destructive, and high-precision detection of internal diseases in apples.