Smart Threshold: X-ray CT adapts apple segmentation across 12 cultivars

Every apple on a supermarket shelf carries two invisible histories: the genetics of its cultivar and the orchard management that shaped its growth. Apple fruit quality is jointly determined by cultivar genetics and preharvest management, yet most non-destructive assessment methods treat these sources of variation as inseparable. Breeding programmes select for cultivar-specific traits while postharvest management targets environmental predictors, and this separation limits quantitative understanding of how structural phenotypes respond to each factor. Without tools that characterise internal fruit architecture across both genetic backgrounds and management conditions, cross-cultivar quality standards remain difficult to establish. A new study published in Smart Agricultural Technology tackles this problem head-on, using X-ray computed tomography to peer inside intact apples and profile their internal architecture across twelve genetically diverse cultivars and a controlled crop-load experiment within a single analytical framework.

X-ray computed tomography enables non-destructive extraction of internal structural features from intact fruit. Applications to apple and pear tissue have demonstrated porosity mapping, cellular-level characterisation, and automated disorder detection via deep learning. However, most CT-based quality studies apply fixed segmentation parameters calibrated on a single cultivar or narrow cultivar range. For example, one study applied a fixed porosity threshold optimised for Fuji apples, while another used uniform grey-value thresholds across bruise detection tasks without cultivar-specific calibration. When tissue density varies substantially across cultivars, such fixed thresholds produce systematic segmentation errors that propagate into all downstream features. This limitation restricts the applicability of CT-derived quality metrics across genetically diverse germplasm. A second gap concerns the interpretation of CT features across experimental contexts: cultivar classification studies have used CT-derived features to distinguish varieties based on shape or internal structure, but without assessing whether the same features also respond to management variation. Conversely, investigations of crop-load or storage effects have quantified structural changes within a single cultivar without evaluating cross-cultivar generalisability. To the authors’ knowledge, no prior work has profiled the same CT feature set across both a multi-cultivar diversity panel and a controlled management trial within a single framework.

The research team, led by Weiqun Wang of McGill University together with Dario Mengoli, Shangpeng Sun, Luigi Manfrini, and Shirin Ghatrehsamani, designed a dual-batch experiment to disentangle these factors. Batch 1 comprised 36 apple fruit from 12 cultivars (three replicates each) grown under identical conditions at the McGill University Macdonald Campus orchard. The cultivars spanned commercial and heritage germplasm: Altas, Ambrosia, Cortland, Empire, Gala, Ginger Gold, Honey Crisp, Liberty, Lobo, McIntosh, Nova Easygro, and Nova Spy. Because all samples shared the same orchard environment, structural differences among cultivars are attributed to genetic background. Batch 2 comprised 27 Gala fruit from a 3 × 3 factorial design with three crop-load levels (high 92 fruits per tree, standard 46, low 23) and three harvest dates spanning 100 to 134 days after full bloom, with three replicates per treatment combination. Crop-load treatments were applied by manual thinning in early July at 20 to 30 mm fruit diameter. Because genetic background was held constant, structural differences are attributed to management-induced physiological variation. All fruit were stored at 4 °C and 40 percent relative humidity pending scanning.

Volumetric scanning was performed on a Canon Aquilion Prime SP system at 80 kV and 50 mA, yielding 339 µm isotropic voxel resolution. Each 10 to 13 second scan generated approximately 1,200 projection images in DICOM format at 0.5 mm inter-slice thickness. A total of 21 sessions acquired complete volumetric data for all 63 samples. DICOM volumes were reconstructed in Python using NumPy and SciPy. Individual fruit were segmented from background by adaptive intensity thresholding followed by three-dimensional morphological operations. Internal structures including seeds, cavities, skin, flesh, and core were classified via intensity-based detection with spatial topology constraints. Tissue porosity was calibrated by linear regression between two reference standards: Gala apple juice (grey value 49, 0 percent porosity) and external air (grey value −1007, 100 percent porosity), yielding an R² of 0.957. Twenty-five features were extracted per sample in three categories: morphological (total volume, flesh volume, core volume, seed count), density (grey-value statistics across flesh, core, and skin compartments), and porosity (whole-fruit, flesh, and core porosity). All features were z-score normalised, and only features measurable in both batches were retained for cross-batch comparison.

The centrepiece of the study is an adaptive flesh-to-core segmentation method the authors call Smart Threshold. Fixed porosity thresholds fail when tissue density varies across cultivars. To address this, the team developed an approach that generates a radial porosity profile from the equatorial cross-section of each sample and smooths it with a Gaussian filter. The core boundary is placed at the inflection point: the minimum radial porosity plus a 0.7 percent offset. This value was selected based on the typical porosity gradient observed at the seed-cavity transition in preliminary scans of structurally contrasting cultivars, where offsets below 0.5 percent failed to exclude seed-adjacent voxels and offsets above 1.5 percent encroached into flesh tissue. Sensitivity analysis across offsets of 0.3 to 2.0 percent confirmed that cultivar discrimination and classification accuracy were invariant to this parameter, indicating that downstream results are robust to the specific offset chosen. The resulting threshold varied from 10.1 percent in Gala to 22.6 percent in Lobo across the 12 cultivars, compared with a fixed threshold of 20 percent that would produce segmentation errors of up to 77 percent relative deviation in dense-tissue cultivars. Validity was confirmed by morphological consistency with CT cross-sections across all 12 cultivars, with core boundaries aligning to visually identifiable seed-cavity transitions in every case.

Adaptive segmentation revealed substantial variation in internal fruit architecture across the 12 cultivars grown under identical conditions. Flesh content ranged from 59.7 percent in Gala and Nova Spy to 85.7 percent in Lobo, a 26-percentage-point span. Flesh porosity (19.1 to 32.8 percent) and core porosity (9.9 to 19.2 percent) varied independently across cultivars. Flesh percentage correlated positively with flesh porosity, indicating that cultivars allocating more tissue to flesh also develop greater intercellular air space. Three descriptive structural groupings emerged from these profiles. Size-prioritising cultivars such as Lobo and Nova Easygro maximised flesh proportion above 74 percent with moderate core porosity. Architecture-balanced cultivars such as Altas and McIntosh maintained intermediate flesh content around 70 percent with simultaneously high flesh porosity above 30 percent and core porosity above 18 percent, a combination absent in other cultivars. Dense-tissue cultivars such as Gala and Honey Crisp exhibited consistently low flesh porosity below 22 percent and the highest core proportions above 29 percent. These strategies are reflected in the adaptive threshold values: dense-tissue cultivars required thresholds near 11 percent, whereas size-prioritising cultivars required thresholds above 20 percent.

Random Forest classification with leave-one-out cross-validation achieved 72.2 percent internal cross-validation accuracy across 12 cultivars, exceeding the 8.3 percent random baseline. Given only three replicates per cultivar, this internal estimate should not be interpreted as external classification performance. The confusion matrix showed that architecture-balanced cultivars reached above 85 percent per-class accuracy under leave-one-out validation, while structurally similar cultivars within the same descriptive group exhibited higher mutual confusion. Together, these results are consistent with CT features carrying cultivar-associated structural information within this dataset. The adaptive threshold correlated strongly with core porosity (r = 0.990) and moderately with flesh porosity (r = 0.841) across all 63 samples from both batches, consistent with the threshold capturing an anatomically interpretable tissue-density gradient rather than an arbitrary segmentation artefact. Within Batch 2, the adaptive threshold responded significantly to crop-load treatment, with low crop-load fruit exhibiting higher thresholds than high crop-load fruit, demonstrating that the same segmentation parameter that distinguishes cultivars also captures management-induced structural modification within a single cultivar.

Gaussian Mixture Model clustering of the two-dimensional signal space, plotting genetic signal from Random Forest importance against physiological signal from Kruskal-Wallis significance, identified four feature categories. Genetic-dominant features, comprising 32 percent of the set, were dominated by density metrics such as grey-value mean, standard deviation, and flesh and whole-fruit porosity that varied strongly across cultivars but showed minimal crop-load sensitivity. Physiology-dominant features, comprising 28 percent, were dominated by size-related metrics such as total volume, flesh volume, core volume, and centre coordinates responsive to crop-load treatment. Dual-sensitive features, comprising 20 percent, showed elevated signal in both dimensions: Smart Threshold, GV Median, GV Core Mean, Porosity Core, and Skin Volume showed significance in both batches. Four of these five features were highly inter-correlated, indicating that the effective dimensionality of the dual-sensitive signal is lower than the feature count. Noise or weak features, comprising the remaining 20 percent, showed minimal signal in either dimension. The dual-sensitive features all exhibited large effect sizes, with Cohen’s d values ranging from 0.70 to 2.10, and six additional features previously reported as crop-load-responsive were confirmed with comparable effect sizes.

Principal component analysis showed that seven components captured 90 percent and nine components captured 95 percent of cumulative variance, indicating that the 25-feature set contains approximately seven to nine independent dimensions of structural information. The authors frame their findings with appropriate caution. Batch 2 used a single cultivar at one site, so genotype-by-environment interactions could alter which features qualify as dual-sensitive under different cultivar-environment combinations. No quantitative boundary metrics such as Dice coefficients against manual annotation were computed, constituting a methodological gap. The small per-cultivar sample size constrains the precision of cultivar-level estimates and the statistical power of the 12-class classification. The study assessed static structural properties at harvest, and developmental sampling across multiple growth stages combined with gene expression data would clarify the temporal sequence of genetic specification and environmental modification. The dual-sensitive features identified here represent candidate structural markers, not validated quality indicators, and any application to postharvest grading or precision quality assessment remains explicitly conditional until tested against functional quality outcomes such as firmness, mealiness, and storage disorders.

The study is best framed as a methodological framework and a set of candidate features that warrant broader validation across cultivar diversity, growing environments, and functional quality endpoints. Cross-cultivar validation across diverse genetic backgrounds, production regions, and seasons is needed to quantify genotype-by-environment interactions and test the stability of the feature sensitivity patterns reported here. Validation against functional quality parameters such as firmness, sugar content, and storage disorder incidence is required before structural markers can be recommended for postharvest use. The development of non-destructive proxies for dual-sensitive features would enable practical field application without CT infrastructure. The adaptive segmentation framework is conceptually applicable to other pome and stone fruits, though empirical validation in those systems would be necessary. For now, the work demonstrates that cultivar-dependent segmentation thresholds can substantially affect CT-derived structural measurements in apple fruit, and that adaptive segmentation offers a potentially useful strategy for reducing such discrepancies across the 12 cultivars and single site examined here.

Subject of Research: Adaptive X-ray CT segmentation (Smart Threshold) profiles apple fruit internal structure across 12 cultivars and a Gala crop-load trial, identifying dual-sensitive structural features.

Article Title: Cultivar-adaptive X-ray CT tissue segmentation and multi-scale structural feature profiling in apple fruit

Article References: Wang, W., Mengoli, D., Sun, S., Manfrini, L., & Ghatrehsamani, S. (2026). Cultivar-adaptive X-ray CT tissue segmentation and multi-scale structural feature profiling in apple fruit. Smart Agricultural Technology, 15, Article 102624. https://doi.org/10.1016/j.atech.2026.102624

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102624

Keywords: apple, X-ray computed tomography, fruit quality, segmentation, machine learning, phenotyping, Cultivar-adaptive, X-ray, tissue, multi-scale, structural, feature