Simultaneous non-destructive detection of watercore and bruising in ‘Fuji’ apples via X-Ray computed tomography: Dual-threshold segmentation and machine learning classification

Abstract

In markets such as China and Japan, watercore and bruising are decisive factors determining the commercial value of premium-grade 'Fuji' apples. This study presents a non-destructive approach integrating X-ray computed tomography (CT) with image processing and machine learning (ML) to simultaneously detect watercore and bruising, and grade watercore severity. A threshold-based segmentation protocol (grayscale value, GSV > 437 for watercore; GSV < 0 for bruising) enabled precise 3D reconstruction of affected tissues. Key morphological and statistical parameters were extracted, with volume3d, average GSV, and perimeter exhibiting strong correlations to watercore index (WI) (R > 0.8, p < 0.001). Three ML classifiers—LDA, SVM, and RF—achieved validation accuracies of 82.87 %, 91.20 %, and 90.28 % for four-grade WI classification (Normal, WI=1–3), supported by high AUCs (>0.9). SHAP analysis confirmed cross-model consistency in feature importance (volume3d, A-GSV, perimeter). Crucially, CT tracked watercore remission during storage: watercore volume fraction (WVF) declined from 13.04 % (WI=1), 17.53 % (WI=2), and 33.5 % (WI=3) to 3.6–10.1 % after 9 weeks at 4°C. We established a diagnostic WVF threshold of 11.2 % and proposed storage protocols: WI= 1 for immediate sale, WI= 2 storage ≤ 6 weeks, WI= 3 consumption ≤ 8 weeks. The framework optimizes sensory quality, market value, and supply chain efficiency, providing a scientific basis for the development of targeted distribution strategies.