Microplastics (MPs) contamination has emerged as a pervasive environmental and public health concern, with MPs detected in marine ecosystems, freshwater systems, sediments, soils, and even in food items such as seafood, beer, canned sardines, honey, and drinking water. These findings highlight the inevitability of human exposure through ingestion, inhalation, and dermal contact, with recent estimates suggesting an annual intake of 39,000 - 52,000 plastic particles per person, likely underestimated when considering airborne deposition during meals. Conventional analytical techniques for MPs detection in biological matrices, such as pyrolysis-GC/MS and spectroscopic methods (FTIR, Raman), are widely used but suffer from critical limitations: they require destructive sample preparation, are highly time-consuming (often 3 - 4 days per sample), and fail to preserve spatial information regarding MPs localization within tissues. These constraints hinder comprehensive toxicological assessments and increase the risk of cross-contamination during processing. Consequently, there is an urgent need for non-destructive, rapid, and reliable methodologies capable of detecting MPs in complex biological samples while maintaining positional integrity. Recent advances in imaging and artificial intelligence (AI) offer promising solutions to these challenges. Computed tomography (CT) and microCT have demonstrated potential for non-invasive MPs detection in sediments and soils; however, their application to biological samples has been limited until recently. MicroCT has recently emerged as a novel method for detecting MPs. It was first applied in commercial fish and later in zebrafish, allowing researchers to obtain three-dimensional images of MPs without the need for destructive sample preparation. Importantly, microCT preserved tissue integrity, minimized contamination risk, and allowed for spatial mapping of MPs accumulation sites, critical for correlating MPs presence with potential toxicological effects. Building on these developments, an automated CT-based methodology has been proposed, combined with deep learning for MPs detection in fish samples. This approach demonstrated several advantages over traditional techniques: accurate localization of MPs, reduced contamination risk, rapid processing (a few hours versus several days), and scalability to larger sample volumes (up to 100 cm³) at a relatively low cost (~€150 per sample). The semantic segmentation model achieved near-perfect detection of inoculated MPs, highlighting the potential of AI-driven automation to overcome the limitations of manual spectral analysis. Similarly, the importance of uncertainty quantification in CT and neural network workflows 5 has been emphasized, identifying key sources of error, such as voxel resolution, image noise, and algorithmic variability, and proposing adaptive thresholding and calibration strategies to mitigate false positives and false negatives. Their findings underscore that robust uncertainty management is essential for ensuring reliability in AI-assisted MPs detection. The integration of AI with CT imaging represents a transformative step in MPs research. Unlike FTIR and Raman spectroscopy, which require extensive sample preparation and are sensitive to additives or surface modifications, deep learning-based segmentation can leverage multifactorial information (density, shape, spatial context) to accurately identify MPs without destructive processing. This automation significantly reduces operator dependency and enables high-throughput analysis, paving the way for real-time monitoring applications in food safety and environmental assessment. Furthermore, the proposed methodology offers a favorable cost-benefit ratio, combining rapid acquisition and automated segmentation with minimal labor requirements, thereby addressing critical bottlenecks in conventional workflows. In conclusion, the convergence of CT imaging and AI-driven segmentation offers a robust, non-destructive, and cost-effective solution for MPs detection in biological samples. By enabling rapid, automated, and spatially resolved analysis, this methodology addresses critical limitations of conventional workflows and opens new avenues for toxicological studies, risk assessment, and food quality control. Continued efforts to refine resolution limits, manage uncertainty, and broaden applicability across diverse matrices will be essential to fully realize its potential in environmental and health research. This integrated approach not only accelerates MPs detection but also provides a foundation for future innovations aimed at mitigating the ecological and health impacts of plastic pollution.
La contaminazione da microplastiche (MPs) è emersa come una preoccupazione ambientale e sanitaria sempre più diffusa, con MPs rilevate negli ecosistemi marini, nei sistemi di acqua dolce, nei sedimenti, nei suoli e persino in alimenti come frutti di mare, birra, sardine in scatola, miele e acqua potabile. Questi risultati evidenziano l’inevitabilità dell’esposizione umana attraverso ingestione, inalazione e contatto dermico, con stime recenti che suggeriscono un’assunzione annuale di 39.000–52.000 particelle di plastica per persona, probabilmente sottostimata se si considera la deposizione aerea durante i pasti. Le tecniche analitiche convenzionali per il rilevamento delle MPs in matrici biologiche, come la pirolisi-GC/MS e i metodi spettroscopici (FTIR, Raman), sono ampiamente utilizzate ma presentano limiti critici: richiedono preparazioni distruttive del campione, sono estremamente dispendiose in termini di tempo (spesso 3–4 giorni per campione) e non preservano le informazioni spaziali sulla localizzazione delle MPs nei tessuti. Questi vincoli ostacolano valutazioni tossicologiche complete e aumentano il rischio di contaminazione crociata durante la lavorazione. Di conseguenza, è urgente disporre di metodologie non distruttive, rapide e affidabili, in grado di rilevare MPs in campioni biologici complessi mantenendo l’integrità posizionale. I recenti progressi nell’imaging e nell’intelligenza artificiale (AI) offrono soluzioni promettenti a queste sfide. La tomografia computerizzata (CT) e la microCT hanno mostrato potenziale per il rilevamento non invasivo delle MPs in sedimenti e suoli; tuttavia, la loro applicazione ai campioni biologici è stata limitata fino a poco tempo fa. La microCT è recentemente emersa come un nuovo metodo per il rilevamento delle MPs: è stata applicata per la prima volta in pesci commerciali e successivamente nello zebrafish, consentendo di ottenere immagini tridimensionali delle MPs senza necessità di preparazioni distruttive. In modo significativo, la microCT ha preservato l’integrità dei tessuti, ridotto il rischio di contaminazione e permesso la mappatura spaziale dei siti di accumulo, fondamentale per correlare la presenza di MPs con potenziali effetti tossicologici. Sulla base di questi sviluppi, è stata proposta una metodologia automatizzata basata su CT, combinata con deep learning, per il rilevamento delle MPs nei campioni ittici. Questo approccio ha mostrato diversi vantaggi rispetto alle tecniche tradizionali: localizzazione accurata delle MPs, riduzione del rischio di contaminazione, rapidità di elaborazione (poche ore contro diversi giorni) e scalabilità a volumi di campione maggiori (fino a 100 cm³) a un costo relativamente basso (~150 € per campione). Il modello di segmentazione semantica ha ottenuto un rilevamento quasi perfetto delle MPs inoculate, evidenziando il potenziale dell’automazione basata su AI nel superare i limiti dell’analisi spettrale manuale. Allo stesso modo, è stata sottolineata l’importanza della quantificazione dell’incertezza nei flussi di lavoro CT e reti neurali, identificando fonti chiave di errore come la risoluzione dei voxel, il rumore dell’immagine e la variabilità algoritmica, e proponendo strategie di soglia adattiva e calibrazione per mitigare falsi positivi e falsi negativi. Questi risultati mostrano che una gestione robusta dell’incertezza è essenziale per garantire l’affidabilità nel rilevamento delle MPs assistito da AI. L’integrazione dell’AI con l’imaging CT rappresenta un passo trasformativo nella ricerca sulle MPs. A differenza della spettroscopia FTIR e Raman, che richiede preparazioni estese e risulta sensibile ad additivi o modifiche superficiali, la segmentazione basata su deep learning può sfruttare informazioni multifattoriali (densità, forma, contesto spaziale) per identificare accuratamente le MPs senza processi distruttivi. Questa automazione riduce significativamente la dipendenza dall’operatore e consente analisi ad alto throughput, aprendo la strada a applicazioni di monitoraggio in tempo reale nella sicurezza alimentare e nella valutazione ambientale. Inoltre, la metodologia proposta offre un rapporto costo-beneficio favorevole, combinando acquisizione rapida e segmentazione automatizzata con un fabbisogno minimo di manodopera, affrontando così i principali colli di bottiglia dei metodi convenzionali. In conclusione, la convergenza tra imaging CT e segmentazione guidata da AI offre una soluzione robusta, non distruttiva ed economicamente vantaggiosa per il rilevamento delle MPs nei campioni biologici. Consentendo analisi rapide, automatizzate e spazialmente risolte, questa metodologia supera i limiti critici dei flussi di lavoro tradizionali e apre nuove prospettive per studi tossicologici, valutazioni del rischio e controllo della qualità alimentare. Sarà essenziale continuare a migliorare i limiti di risoluzione, gestire l’incertezza e ampliare l’applicabilità a matrici diverse per sfruttare appieno il potenziale di questo approccio nella ricerca ambientale e sanitaria. Questa strategia integrata non solo accelera il rilevamento delle MPs, ma fornisce anche una base per future innovazioni volte a mitigare gli impatti ecologici e sanitari dell’inquinamento da plastica.
DETECTION OF MICROPLASTICS IN FISH USING COMPUTED TOMOGRAPHY AND DEEP LEARNING / Strafella, Pierluigi. - (2026 Mar).
DETECTION OF MICROPLASTICS IN FISH USING COMPUTED TOMOGRAPHY AND DEEP LEARNING
STRAFELLA, PIERLUIGI
2026-03-01
Abstract
Microplastics (MPs) contamination has emerged as a pervasive environmental and public health concern, with MPs detected in marine ecosystems, freshwater systems, sediments, soils, and even in food items such as seafood, beer, canned sardines, honey, and drinking water. These findings highlight the inevitability of human exposure through ingestion, inhalation, and dermal contact, with recent estimates suggesting an annual intake of 39,000 - 52,000 plastic particles per person, likely underestimated when considering airborne deposition during meals. Conventional analytical techniques for MPs detection in biological matrices, such as pyrolysis-GC/MS and spectroscopic methods (FTIR, Raman), are widely used but suffer from critical limitations: they require destructive sample preparation, are highly time-consuming (often 3 - 4 days per sample), and fail to preserve spatial information regarding MPs localization within tissues. These constraints hinder comprehensive toxicological assessments and increase the risk of cross-contamination during processing. Consequently, there is an urgent need for non-destructive, rapid, and reliable methodologies capable of detecting MPs in complex biological samples while maintaining positional integrity. Recent advances in imaging and artificial intelligence (AI) offer promising solutions to these challenges. Computed tomography (CT) and microCT have demonstrated potential for non-invasive MPs detection in sediments and soils; however, their application to biological samples has been limited until recently. MicroCT has recently emerged as a novel method for detecting MPs. It was first applied in commercial fish and later in zebrafish, allowing researchers to obtain three-dimensional images of MPs without the need for destructive sample preparation. Importantly, microCT preserved tissue integrity, minimized contamination risk, and allowed for spatial mapping of MPs accumulation sites, critical for correlating MPs presence with potential toxicological effects. Building on these developments, an automated CT-based methodology has been proposed, combined with deep learning for MPs detection in fish samples. This approach demonstrated several advantages over traditional techniques: accurate localization of MPs, reduced contamination risk, rapid processing (a few hours versus several days), and scalability to larger sample volumes (up to 100 cm³) at a relatively low cost (~€150 per sample). The semantic segmentation model achieved near-perfect detection of inoculated MPs, highlighting the potential of AI-driven automation to overcome the limitations of manual spectral analysis. Similarly, the importance of uncertainty quantification in CT and neural network workflows 5 has been emphasized, identifying key sources of error, such as voxel resolution, image noise, and algorithmic variability, and proposing adaptive thresholding and calibration strategies to mitigate false positives and false negatives. Their findings underscore that robust uncertainty management is essential for ensuring reliability in AI-assisted MPs detection. The integration of AI with CT imaging represents a transformative step in MPs research. Unlike FTIR and Raman spectroscopy, which require extensive sample preparation and are sensitive to additives or surface modifications, deep learning-based segmentation can leverage multifactorial information (density, shape, spatial context) to accurately identify MPs without destructive processing. This automation significantly reduces operator dependency and enables high-throughput analysis, paving the way for real-time monitoring applications in food safety and environmental assessment. Furthermore, the proposed methodology offers a favorable cost-benefit ratio, combining rapid acquisition and automated segmentation with minimal labor requirements, thereby addressing critical bottlenecks in conventional workflows. In conclusion, the convergence of CT imaging and AI-driven segmentation offers a robust, non-destructive, and cost-effective solution for MPs detection in biological samples. By enabling rapid, automated, and spatially resolved analysis, this methodology addresses critical limitations of conventional workflows and opens new avenues for toxicological studies, risk assessment, and food quality control. Continued efforts to refine resolution limits, manage uncertainty, and broaden applicability across diverse matrices will be essential to fully realize its potential in environmental and health research. This integrated approach not only accelerates MPs detection but also provides a foundation for future innovations aimed at mitigating the ecological and health impacts of plastic pollution.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


