CNN

This page describes the CNN row of the Papers.md matrix: convolutional neural networks for grid- or image-structured data, covering microscopy and product imaging, hyperspectral and spectral data, and 3D scaffold tomography. The row’s authoritative scope is its Taxonomy.md definition; this page synthesizes what currently sits in it.

Scope boundary

The taxonomy sets a substantive bar: the placement requires genuine spatial convolution on image-like input. A dense regression surrogate labelled “CNN” in its own paper belongs in Deep Learning instead. That test is worth applying carefully, because “CNN” is used loosely in the food-science literature, and because a 1D convolution over a spectrum genuinely does qualify: the wavelength axis is a grid and local structure along it is real.

The row is the most physically grounded on the method axis. Its inputs are things a cell-ag facility already produces, brightfield images of a culture, hyperspectral scans of a reactor or a product, confocal images of a gel, tomography of a scaffold, and its outputs are measurements someone currently gets by destroying a sample or waiting weeks for a test. Several of these papers are best read as replacements for an assay rather than as models.

Cellular Engineering

#218 (Yang et al. 2025, Food Research International): scores adipogenic differentiation of porcine cells in suspension from brightfield images, replacing fluorescence staining and molecular assays that are slow, laborious and observer-dependent. Fat is a primary determinant of meat quality, so this is a directly product-relevant readout, and the culture is a serum-free suspension process using commercial cultivated-meat cell lines and media rather than a research proxy. One of the small number of matrix references working on a real cultivated-meat production system.
#264 (Sinha et al. 2025, Frontiers in Artificial Intelligence, Michigan State): predicts fluorescent markers directly from brightfield microscopy of bovine satellite cells, a U-Net predicting DAPI and Pax7 signals simultaneously as proxies for cell abundance and differentiation status. This is digital staining: the point is not accuracy for its own sake but that the sample survives, so the same culture can be assessed repeatedly rather than fixed and discarded. Cells were isolated from Longissimus muscle of three-month-old Holstein bull calves, with a preprocessing pipeline including fluorescent-signal denoising. Code at food-ai-engineering-lab/bsc-fluorescence-prediction.
#265 (Nabiullina et al. 2025, Frontiers in Bioscience-Landmark): pairs 3D bioprinting of a hybrid meat product with a fine-tuned YOLO model detecting and counting lipoblasts, fibroblasts and myogenic cells. Mesenchymal stem cells and fibroblasts came from rabbit, chosen because it is both a meat species and a research model, obtained by laparoscopic biopsy so the donor survives, which the authors present as consistent with the cultivated-meat premise. Cell lines were STR-validated and mycoplasma-tested, a QC step worth noting because it is often absent.

Bioprocess & Scale-Up

#206 (Hevaganinge et al. 2023, ACS Omega): quantifies glucose and lactate in CHO culture from contactless short-wave infrared hyperspectral imaging, addressing a specific failure of existing spectral sensors, that they must be submerged and then foul with cell debris. The architecture is designed for interpretability rather than raw accuracy: a convolutional encoder’s latent space is used first for sparse regression onto metabolite concentrations, then the learned weights select the best convolution filters to feed parallel interpretable models such as Gaussian process regression, and the standard deviation across predictions per image automatically filters out spurious ones. Benchmarked against polynomial PLS on mean spectra. The self-filtering step is what makes it plausible as feedback control rather than monitoring.
#260 (Iseoka et al. 2022, Journal of Artificial Organs): a CNN-based mycoplasma detection system, aimed squarely at manufacturing cost. The conventional mycoplasma test takes several weeks, is expensive, and requires expert reading of images. The program has three parts, detection, prediction and cell counting, and is built so a user can verify which cells it identified as infected rather than accept a verdict. For cultivated meat the same arithmetic holds: contamination testing is a release gate, and its turnaround time bounds how fast a facility can run.
#261 (Maruthamuthu et al. 2020, MicrobiologyOpen): detects microbial contamination from Raman spectra with a CNN, over a purpose-built dataset of 12 microbes spanning Gram-positive, Gram-negative and fungal contaminants, chosen from the FDA’s manual of pharmaceutical microbiology plus environmental sources, measured alone and mixed with CHO cells. Reported 95% to 100% classification accuracy. Together with #259 and #262 in SVM, this is a small but coherent body of work on the same problem by three different routes.
#33 (Rojek et al. 2021, ICCS): accelerates CFD simulation of chemical mixing by pairing an OpenFOAM-based solver with a CNN surrogate. The design choice worth carrying is the decomposition: a separate sub-model is trained per simulated quantity, so at inference they can be pipelined to cut execution time or called one at a time to cut memory. Mixing is the physics that governs bioreactor scale-up, and the paper’s framing (that building a bespoke simulator costs years and trades generality for accuracy in a narrow range) is the argument for surrogates generally. Catalogued alongside OpenFOAM.

Scaffolding

#19 (Bermejillo Barrera et al. 2021, Materials): 3D CNNs over virtual tomography of tissue-engineering scaffolds, predicting mechanical properties from the voxelized geometry. A library of 20 lattice geometries, all inscribed in a 5 × 5 × 5 mm cube and deliberately varied between periodic, symmetric designs and irregular truss interconnections, is characterized by CAD-measured porosity and finite-element simulation. The motivation is that conventional design and simulation fail on some scaffolds outright, through geometric complexity, manufacturing defects, or large aspect ratios causing numerical mismatch, which is when a learned surrogate stops being a convenience and becomes the only route.
#310 (Kim et al. 2026, Current Research in Food Science): predicts protein composition and hardness of 3D-printed plant-based meat analogues from RGB surface images captured during printing, using DeepLabv3+ with a ResNet-50 backbone to semantically segment printability features (pores, aggregation, spreading). Twenty-two formulations varying wheat gluten, soy protein isolate and rice protein were each printed in triplicate under a three-component design. Reported mIoU 0.558 and mAcc 0.620 for segmentation. The interesting claim is the chain: surface appearance during printing carries enough information to infer both the formulation and a mechanical property, non-destructively and in process.
#349 (Yang 2026, Colloids and Surfaces A): benchmarks eight CNN architectures for segmenting protein-gel networks in confocal microscopy, spanning different decoders, encoder backbones and attention mechanisms (UNet-R34, DeepLabv3+-R50, FPN-R34, LinkNet-R34, MANet-R34, UNet-EffB0, UNet++-scSE, UNet-scSE). Its finding is a methodological warning: most architectures reach comparable pixel-level accuracy, which is why the paper goes on to evaluate the topology-derived descriptors that actually matter for functionality. Ground truth was built by standardized Fiji thresholding then manual refinement by two experts, with deliberate care to preserve thin interconnecting strands and loops because those dominate the topological measures. Also in Benchmarks & Evaluation Frameworks. Code at yangzhi1987-sketch/benchmark-food-gel-segmentation.

Sensory Prediction

#26 (Sun et al. 2023, Food Chemistry): off-flavour profiling of cultured salmonids from NIR hyperspectral imaging, with a 1D CNN among four classifiers over 107 fish. Described in SVM.
#11 (Shen et al. 2024, Food Chemistry): salted-goose breed discrimination, where a CNN is one of the base classifiers under the GAN-augmented ensemble. Described in GAN / VAE.
#195 (Ulucan et al. 2019, ASYU): monitors a tray of red-meat cubes with a fixed camera over time and classifies quality from RGB images, aimed at retail counters where deterioration means disposal, with the goals stated as reducing waste and keeping open-counter sales quality high. A short conference paper; read the source before relying on architectural detail, since the published description is brief.
#196 MeatScan (Gyening et al. 2025, Data in Brief): a dataset paper rather than a model, contributing 11,000 high-resolution RGB images of cow meat labelled fresh or spoiled, collected across Ghanaian open-air markets, butcher shops and cold rooms at different times of day specifically to capture real variation in lighting, background and moisture. Labels rest on visible spoilage cues (discolouration, sliminess, mould, dryness) verified by trained collectors under natural light. Data at Zenodo. It is the kind of resource the row is short of: most image datasets here are laboratory-clean.
#321 (Ameta et al. 2024, Scientific Reports): predicts odour from vibrational spectra, and does so by turning the spectra into images. Gramian Angular Fields and Markov Transition Fields convert each 1D spectrum into a 2D representation so ordinary computer-vision architectures apply, over 3,018 simulated spectra and a combined odour dataset of 7,374 molecules across 109 classes drawn from the Firmenich “Learning to Smell” challenge and Leffingwell PMP 2001. The framing is a genuine scientific question rather than a benchmark: the vibrational theory of olfaction has gained ground but lacks validation, and the paper tests it with classification, clustering and explainable-AI saliency.
#330 (Sharma et al. 2021, Journal of Chemical Information and Modeling): predicts smell from structure via two parallel representations, physicochemical properties plus fingerprints feeding random-forest and dense models, and 2D chemical images feeding a CNN baseline plus transfer learning from VGG16 and Xception. The dataset was web-scraped from PubChem, TGSC, Sigma-Aldrich, FlavorBase, Flavornet, SuperScent, OdorRactor and AromaDB, with 2,173 of 6,855 entries removed as overlapping or redundant, leaving 4,682 odorants and 542 percepts. That deduplication ratio is worth remembering when reading any odour-prediction result: a third of the naive corpus was duplicate. Also in Deep Learning.

Adjacent methods

Deep Learning: the catch-all for networks without genuine spatial convolution, including dense surrogates that call themselves CNNs.
Chemometrics and SVM: the classical baselines every hyperspectral paper here compares against, and often does not beat by much.
GAN / VAE: the generative counterpart, including the pix2pix scaffolding surrogate and the InfoGAN augmentation in #11.
Ensemble Learning: the fusion layer over CNN and non-CNN base classifiers.
Benchmarks & Evaluation Frameworks: where #349’s architecture benchmark also sits.

CNN