Research Thomas De Kerf

Abstract

Histopathology is the cornerstone of cancer diagnosis. To detect essential biomarkers such as hormone receptors, pathologists rely on immunohistochemical (IHC) staining, a costly method that requires additional tissue processing. In contrast, hematoxylin and eosin (H&E) slides are routinely prepared in every pathology laboratory. Virtual staining, where generative AI translates H&E images into synthetic IHC, offers a transformative alternative. It has the potential to reduce costs, preserve tissue, and improve access to advanced diagnostics, particularly in low- and middle-income countries. However, key challenges remain. It is still unclear which imaging modality—autofluorescence, brightfield, Raman or hyperspectral—provides the most reliable basis for virtual staining. Progress is also hindered by the lack of large, diverse, and high-quality datasets. This project aims to close these gaps by creating a comprehensive, multimodal dataset spanning diverse stains and tissue types. At the same time, we will systematically evaluate the strengths and limitations of each modality in terms of acquisition speed, biochemical specificity, and robustness in clinical workflows. By doing so, we will lay the groundwork for robust and generalizable AI models in digital pathology. The project directly aligns with European health research priorities on affordable, accessible and data-driven innovation. It paves the way for a new generation of diagnostic tools that can enhance the quality and equity of cancer care worldwide.

Researcher(s)

• Promoter: Vanlanduit Steve
• Co-promoter: De Kerf Thomas

Research team(s)

• Industrial Vision Lab (InViLab)

Project type(s)

• Research Project

Abstract

Histological analysis is vital in pathology but relies on chemical staining processes that are time-consuming, costly, destructive to samples, and environmentally harmful. Current virtual staining methods attempt to replicate these stains computationally but often use deep learning techniques that can introduce artifacts ("hallucinations"), undermining clinical reliability. Therefore I introduce HyperStain, an innovative approach that combines hyperspectral imaging (HSI) with point-based spectroscopy to generate accurate virtual stains without the drawbacks of chemical staining or deep learning artifacts. By integrating the speed and spatial resolution of HSI with the chemical specificity of spectroscopic techniques like FTIR, or and Raman spectroscopy, we aim to develop an explainable and reliable virtual staining method. The objectives of this project are to (1) develop this novel virtual staining method, (2) identify optimal spectral wavelengths for virtual staining, and (3) validate our approach against existing imaging modalities and deep learning methods. Additionally, we will create and share a large multi-modal, multi-stain virtual staining database to advance research in the field. By eliminating the need for chemical staining and avoiding the pitfalls of black-box AI models, HyperStain offers a faster, cost-effective, and environmentally friendly alternative that preserves tissue samples for further analysis and enhances diagnostic accuracy.

Researcher(s)

• Promoter: Vanlanduit Steve
• Fellow: De Kerf Thomas

Research team(s)

• Industrial Vision Lab (InViLab)

Project type(s)

• Research Project

Abstract

Polyamides are used as technical plastics by numerous medium sized companies in the production for e.g. parts of electric tools, vehicles, or aluminium windows. The complete value chain, from petrochemical via chemical building blocks (e.g. caprolactam) into finished products is strongly represented in the province Antwerp. Today the polyamide value chain is not circular. This project aims to sort polyamides out of complex mixed waste streams, by using chemical and spectral analysis, coupled with artificial intelligence. The polyamide waste stream can than be transformed into new installations designed for chemical recycling, another key industrial activity in Antwerp.

Researcher(s)

• Promoter: Billen Pieter
• Co-promoter: De Kerf Thomas
• Co-promoter: Mercelis Siegfried
• Co-promoter: Vande Velde Christophe
• Co-promoter: Vanlanduit Steve

Research team(s)

• Intelligence in PRocesses, Advanced Catalysts and Solvents (iPRACS)

Project type(s)

• Research Project

Abstract

This project scopes the detection of moisture in historical buildings using hyperspectral imaging technology. Moisture in buildings can originate from wind-driven rain, rising damp, flooding, leaking infrastructure and condensation and periodic changes in moisture content are the main driver for several decay mechanisms. Traditional methods for the detection of moisture, like gravimetric or electrical approaches, are typically labour intensive, invasive and have limited coverage. The development of a hyperspectral scanning system for in-situ applications will allow the detection of anomalies in large-scale structures like buildings. Such anomalies include the presence of water which results in a specific absorption range in the short wave infrared spectrum. The application will capitalize on recent advances in spectral unmixing, to estimate the moisture content of several porous media, including natural stone and brick. The developments will be validated in case studies on pilot sites. This will fundamentally change the methodology of building conservation and restoration, as a more holistic understanding can be developed from whole building images, which will result in more accurate and detailed sampling strategies.

Researcher(s)

• Promoter: De Kock Tim
• Co-promoter: De Kerf Thomas
• Co-promoter: Koirala Bikram
• Fellow: Sunil John

Research team(s)

• Antwerp Cultural Heritage Sciences (ARCHES)

Project type(s)

• Research Project

Abstract

The impending 31% surge in cancer incidences by 2030, coupled with a critical shortage of histopathologists, underscores an urgent need for innovations in diagnostic methodologies. A time-intensive aspect of histopathology is the staining of tissue slices, a pivotal step for disease diagnosis and research. Recently, hyperspectral imaging has been proposed to generate virtual stains on unstained tissues, a technique that could revolutionize tissue analysis. This method promises reduced errors, increased efficiency, multi-staining capabilities, and sample conservation. However, the technique is currently limited by small sample sizes, undefined wavelength band efficacy, and restricted data accessibility. This research project, aims to expand the sample size to 100 slices across four cancer types, employing three different hyperspectral cameras. We will create a comprehensive database, initially using the H&E stain as a reference. The project's second objective is to deploy deep learning algorithms to transform hyperspectral data into virtual stains and to ascertain the most effective wavelength bands. Finally, we aim to share our findings and dataset openly to encourage collaborative advancements. At InVilab, our infrastructure features an extensive array of imaging equipment, including a quantum cascade laser, enhancing our research capabilities in hyperspectral imaging. However, we currently face a shortfall in high-magnification lenses essential for detailed mid-to-long-wave infrared microscopy. An integral component for advancing our research. Securing funding of the BOF SRG will enable the acquisition of these critical lenses. This enhancement is imperative for integrating hyperspectral imaging into clinical practice, offering a strategic solution to the histopathologist shortage and advancing patient care outcomes.

Researcher(s)

• Promoter: De Kerf Thomas

Research team(s)

• Industrial Vision Lab (InViLab)

Project type(s)

• Research Project