‍Key Takeaways:

1. Radiomics as a Tool for Precision Medicine: Radiomics is transforming medical imaging by quantifying thousands of features from lesions and organs, enabling deeper insights into disease progression and treatment response. However, its adoption is still in progress , particularly in clinical settings.
2. Challenges in Standardization: One of the biggest hurdles in radiomics is the variability in image quality across different sites and equipment. Companies like Quibim are working to solve this by developing harmonization algorithms to standardize imaging data, ensuring consistency and comparability in radiomics features.
3. Integration of Deep Learning: The integration of deep learning in radiomics is proving to be a game-changer, allowing for the extraction of deep features that traditional methods cannot capture. This enhances the performance of predictive models, making them more actionable in clinical contexts.

‍

To start, Angel, could you please share a bit about your background and Quibim?

I'm from Spain and hold a PhD in biomedical engineering. I graduated in 2007 with a degree in electronics and computer science, then pursued a PhD in image processing for MRI. Over the years, I've specialized in computer vision applied to medical imaging, focusing on quantifying properties from medical images, a field we termed Imaging Biomarkers. Along with colleagues, we contributed to founding the European Imaging Biomarkers Alliance within the European Society of Radiology. Our mission was to standardize robust methods for quantifying image properties, which are crucial for understanding disease progression and treatment response.

In 2014, the advent of deep learning and GPUs marked a significant shift, allowing for the automation of imaging biomarker extraction. This technological progress enabled us to develop a business model centered around an analysis pipeline—a series of algorithms that process medical images from start to finish, yielding valuable information. For example, a pipeline might analyze a prostate MRI to identify and characterize lesions.

Deep learning significantly reduced the need for manual intervention in these processes, facilitating the development of a business model where patient data could be analyzed to produce reports containing quantitative features or objective findings. To support our model, we founded Quibim, which stands for Quantitative Imaging Biomarkers in Medicine. Our business model involves collaborating with biopharma and academic medical centers to access large patient cohorts for model training, research, and the identification of new biomarkers that can later become a clinical tool, by productizing them as Software as Medical Devices (SaMD).

Quibim has achieved two FDA clearances, one for prostate cancer and another for brain diseases. We are a team of 100 people, headquartered in Spain, with subsidiaries in New York (Quibim Inc.) and the UK (Quibim Ltd.). We have raised around $30 million and collaborate with over 150 hospitals. Our clients include major companies like Novartis, Johnson & Johnson, Merck, Philips, and various hospitals.

Considering radiomics is a term unfamiliar to some, how would you define the word?

Radiomics is a technique that allows for the analysis of characteristics in lesions and organs that are not visible to the naked eye, such as textures, that go beyond shape and size. Unlike basic observations, radiomics quantifies thousands of features from these regions, describing textures, shapes, and sizes through deep statistical analysis on the pixel intensities (voxel, if we think in 3D, given that any MR, PET or CT image represents a tissue slice). This analysis examines the organization of a lesion's inner heterogeneity, providing thousands of quantitative features that describe this organization. Essentially, radiomics transforms an image into thousands of features for statistical processing and analysis.

‍

How would you characterize the current state of radiomics adoption, both within clinical settings and also in research settings?

The current stage of radiomics is such that some tools have already been implemented and cleared as medical devices. For example, we have a prostate cancer classifier based on radiomics features that is now a product and a medical device. However, when considering radiomics as a general term, similar to how one might think of genomics or Next-Generation Sequencing (NGS), which has properties that can be used to stratify patients, radiomics is still in an initial phase. Radiologists are using it to stratify lesions in areas such as the prostate, breast, and rectum, thanks to the development of specific products. However, for some specialists like oncologists, who are the end users in cancer, radiomics is still in an evidence generation phase for them to use it as an actionable tool. It is at an early stage where there are challenges that need to be addressed. This is why I am fully committed to the company that is taking the lead in tackling the most relevant issues concerning radiomics, being the biggest one, harmonization in the image domain (standardizing image quality) and features domain (making sure features are robust enough for predictive/prognostic models).

‍

What are the clinical challenges and barriers you've encountered and what do you anticipate the next wave of radiomics implementation will look like in clinical practice?

Currently, radiomics faces a significant challenge: its lack of actionable insights for accurately stratifying patients for specific treatments or outcomes. Although we have research algorithms that predict overall survival in lung cancer, these are not yet medical devices. To transition them into medical devices, we must conduct multisite, multi-case, multi-reader studies, and pivotal studies to provide the necessary evidence.

One of the most pressing issues in radiomics is evidence generation, which requires the harmonization of imaging data. Currently, radiomics features vary significantly across different hospitals, making comparisons difficult. At Quibim, we've addressed this challenge by standardizing image quality. We created a common reference that ensures images from various manufacturers are adjusted to a uniform dynamic range. This standardization allows for comparable feature extraction from the images.

It's clear that preprocessing images is essential. Relying solely on AI or statistical methods cannot resolve the discrepancies in image quality across sites, machines, and vendors. The variability in image quality is a major obstacle that must be overcome. For example, just as the material used in blood extraction can affect the sample, the imaging equipment used in radiomics can 'contaminate' the images with variability. Our goal is to eliminate this variability by bringing images to a common histogram and dynamic range, ensuring they are comparable.

To achieve this, our company has invested in developing proprietary algorithms for image harmonization. This step is crucial; without harmonization, there is no common language in radiomics. Harmonization is not just beneficial; it's essential for the field's advancement.

‍

As you've mentioned, radiomics relies heavily on imaging data. As these modalities and companies begin to provide higher resolution outputs, how does that change your image preprocessing steps? What challenges do you foresee with data quality, storage, and standardization?

Harmonization algorithms are easy to train. Currently, the variability in medical imaging, such as CT scans and MRI, is mainly due to the physical construction of the scanners and how the images are reconstructed. For example, using less tube current in a CT scanner results in more noise, or different echo times in an MRI can significantly alter the contrast. These physical parameters are the primary factors affecting image quality. However, the evolution of medical imaging technologies is now occurring at the software level, with many using deep learning to reconstruct medical images faster or improve image quality. These techniques allow for faster reconstruction with less data, maintaining quality.

As new types of images are generated, harmonization algorithms can be easily retrained by adding a set of new images to the network and generating a new version. This approach is more efficient than using phantoms—materials sent to different hospitals for scanner calibration—which was logistically challenging.

Harmonization and the extraction of radiomics features could lead to the development of reports for oncologists, similar to current gene panels, indicating a patient's risk level based on imaging data analysis. We already have prototypes of how these panels might look.

‍

How do companies navigate the process of running clinical trials or generating data? What are some of the challenges, and how do you overcome them?

I believe this cannot be accomplished in isolation. We receive thousands of images from various sites and institutions for our ongoing radiomics projects. However, academic medical centers and other research institutions must collaborate to advance radiomics evidence generation. We aim to provide clear answers from a radiomics perspective to the most pressing questions in oncology. If radiomics alone cannot fully explain a clinical problem, we should combine it with other techniques, such as genomics and liquid biopsy. For instance, in lung cancer, the community seeks to identify patients who will respond to immune checkpoint inhibitors or determine which patients are likely to have mutations, guiding treatment decisions. These are common oncology questions that remain unanswered, necessitating collaborative efforts towards evidence generation.

What near-term trends or developments are on the horizon that excite you the most?

I'm mostly excited about the fact that everyone is focusing on analyzing just the tumor through radiomics. However, I believe imaging is so powerful that it allows us to analyze the entire picture. The human body contains a wealth of information. For instance, when analyzing a chest for a lung cancer patient, we observe not just the cancer but also subcutaneous fat, the heart, and calcium in the arteries. This provides a plethora of potentially relevant information for patient outcomes that has yet to be analyzed.

Traditional radiomics approaches extract texture features from pixels using formulas. However, by integrating deep features through a deep learning network trained to recognize images, we can utilize the network's inner features as unsupervised, radiomic deep features. Our experience shows that incorporating these deep learning features into models consistently improves their performance. This indicates that AI can capture information that traditional, handcrafted methods cannot. I believe development will continue in this direction, enhancing models and ultimately making some actionable in a medical context.

‍

Do you have any final comments or anything else you wanted to discuss in today's conversation?

It's crucial to emphasize the importance of doctors becoming familiar with radiomics features. These features, much like the early days of genomics, have unusual names. For instance, terms like CXCR4 or NTRK mutation were once new to the oncology community but became well-known through research. Currently, we are conducting research in radiomics, and I believe it's important to encourage the oncology community to invest in this area. This investment would enable high-quality research to determine when radiomics is the most effective solution and when it is not, similar to how certain techniques may be effective for colon cancer but not for head and neck cancer. Essentially, the goal is to garner interest from the oncology community and biopharma companies to advance in this field.