QUIBIM uses AI to boost
breast cancer research

9 May, 2019

As the first ESMO Breast Cancer Congress unfolded 2-4 May in Berlin, Germany, QUIBIM CEO highlighted the role played by artificial intelligence (AI) in breast cancer research.

The European Society of Medical Oncology focused on the progress in treatment options and improved outcomes for breast cancer (BC) patients during its first meeting dedicated to the topic (¹).

One of the key messages of the conference was that therapeutic innovations should go hand in hand with a multidisciplinary, fully integrated approach to patient care, a scenario in which AI can increasingly help make a difference, according to QUIBIM CEO & founder Angel Alberich-Bayarri.

QUIBIM breast cancer structured report

“AI, and in particular image quantification, can help significantly advance knowledge of this multi-faceted disease, by enabling earlier and better detection,” he said.

From the Pacific to the rest of the world

Recently QUIBIM partnered with the AI Precision Health Institute (AI-PHI) at the University of Hawai‘i Cancer Center in Honolulu, to manage, store and quantitatively analyze medical images and algorithms in breast cancer research.

Using the QUIBIM Precision® image analysis platform, AI-PHI researchers will create large scale imaging repositories with automated extraction of imaging biomarkers to characterize patients’ status. The solution will first be used as the central repository for the mammography studies conducted in the Pacific and will include mammograms from over 5 million women.

The University of Hawai‘i Cancer Center is one of 69 NCI designated cancer centers in the United States and the only one in Hawai‘i and all of the Pacific Islands. It is the only institution that uses AI to analyze medical images to assess health and predict risk of disease in the region.

The project is expected to grow exponentially as its principal investigator Dr. John Shepherd is keen on inviting selected reference centers all across the world to join the network. “It is a very interesting and pioneering work for QUIBIM, and our first big cooperation in the breast cancer setting,” Alberich said.

Seamless integration into clinical workflow

The QUIBIM Precision® image analysis platform enables to incorporate AI algorithms regardless of the institution where it is deployed, to facilitate integration into workflow.

“Researchers can program their own algorithm and code, and add it as a plugin to the platform. This feature, combined with the advanced storage and multi-user annotation capabilities of the platform, allows for a wide adoption within research groups and institutions working with AI. Time for deployment of algorithms in the real world is shortened by 25%, since you have the whole AI pipeline in just one place,” he explained.

The solution, which includes not only quantitative image analysis but also structured reporting capabilities, can be integrated into the radiology workflow and add value to this service. It can for instance be used to integrate AI algorithms to detect cancer without the need of radiologists in a first-read.

Earlier this year the platform received CE Mark certification as class IIa Medical Device, including the imaging biomarker analysis algorithms, the zero footprint DICOM viewer and the platform hosting these components and medical imaging data. The platform is now available for commercial deployment in European hospitals and diagnostic imaging centers.

In addition to the platform, the certification includes 15 algorithms in the five key areas of interest of the company: oncology, neurology, musculoskeletal, liver and lung.

Also included in the CE Mark certification is the DataMiner, which was designed in collaboration with the team of Prof. Daniel Keim from Konstanz University to provide advanced visual analytics of large databases of patients for population health management and scientific exploitation. The DICOM web viewer that enables the user to visualize the images of a sequence, draw ROIs or apply filters is also included in the certification.

QUIBIM applies machine learning to develop tools for imaging data quantification, to accelerate image reconstruction, segmentation, detection and data mining. Beyond the breast, the company has created AI algorithms to help detect changes produced by brain and prostate cancer, but also other diseases in the hematological scenario such as non-Hodgkin’s Lymphoma.

(¹): https://www.esmo.org/Conferences/ESMO-Breast-Cancer-2019?utm_campaign=PRESS%20-%20Breast&utm_source=hs_email&utm_medium=email&utm_content=72075686&_hsenc=p2ANqtz-9pH5mSHFhwzXnTGp6oSizohrD75uXccBxo1woBVGUFhJl3K5y_p4Ms4Hoa3LcwRJ_KvILSmOWhsrZxTZSWLwumEXOnOw&_hsmi=72075686

Quibim’s Strategic Advisor Inés Perea discusses the radical disruptions that quantitative imaging analysis can provide to cancer research

By: Angel Alberich-Bayarri

19 June, 2018

Quibim’s Strategic Advisor Inés Perea discusses the radical disruptions that quantitative imaging analysis can provide to cancer research amongst other areas

Inés Perea is a Doctor that graduated with honors in the Pharmacy and Executive Business Program by Instituto de Empresa of Madrid. Mrs. Perea has extensive and versatile experience spanning more than 20 years in the Pharmaceutical industry working within different roles within the Medical, Access and Commercial (Marketing, Sales, Strategic Planning) areas and international experience in Global and regional teams. Her main area of expertise is oncology and biologics where she has been involved in the strategic planning and commercialization of more than 10 drugs/indications, most of them linked to some form of biomarker diagnosis.

How did you come to know QUIBIM before becoming their strategic advisor?

I was in contact with founder and CEO, Angel Bayarri months prior to my start with Quibim. A mutual friend of ours, a director of a large genetics biomarker laboratory in Spain was working specifically on oncology genetic biomarker testing and I would get involved with his organization in the arenas of strategic value and development, as those are two of my core competencies. I was involved in the development and strategic value of genetic/biologic biomarkers in clinical trials in oncology. It was back in the 90’s when biomarkers were a thing of science fiction. Well, he introduced me to Angel and I really thought he had a fantastic idea when it came to implementation of 3-d imaging biomarkers and using algorithms to quantitatively analyze tumors. Once I learned about Quibim’s technology, I knew that the growth could be exponential if he partnered with strategic partners.

Your field of expertise is Oncology. The Quantitative Medical Imaging field has evolved significantly but we still use RECIST criteria for evaluation of treatment responses, what do you think is the main reason for this? Standardization?

This is the foundation for therapeutic trials. It is a standard when examining patients, however, it has some practical limitations. For example, with the latest breakthrough area in oncology research, immunotherapy, many immuno-oncology drugs are known to cause fluctuations in tumor size, which when analyzed simply with RECIST criteria, would appear as an increase to the tumor size, but it is now understood that tumor fluctuations occur regularly during immunotherapy treatments, often causing the tumor to grow prior to cell DE progression, which is in fact a good sign that would be captured with RECIST criteria well after the fact. In those cases, RECIST does not allow for factoring in the context of the holistic viewpoint that quantitative digital image analysis would facilitate.

What therapeutic indications are best served by incorporating quantitative imaging analysis into clinical research trials?

There is a very simple answer to this question: any therapeutic indication where standardized radiological assessments are a primary or secondary safety or efficacy endpoint of a study. This includes solid tumors, liver diseases, brain diseases, lung diseases, and many others.

Let’s discuss costs. Quantitative imaging analysis sounds expensive. However, automation and extraction of biomarkers as soon as a hospital acquires and uploads images related to a clinical trial can and do have a macroeconomic effect of lowering costs, therefore long-term costs decrease. What do you respond with when someone brings up this objection?

You are right, it is a tremendous cost saver from a macroeconomic point of view. What will continue to help this cause are national health systems, private funding, insurers, and other payers to see the long-term cost savings and efficiencies that quantitative imaging analysis facilitates. Essentially, we need to get quantitative image analysis included in the health system reimbursement system much like traditional biomarkers of a decade ago. As a matter of fact, it took health systems quite a bit of time to include standard biomarkers that we have all grown accustomed to, included on their reimbursement schedules. While it essentially all boils down to payers and health authorities, us researchers and technology pioneers need to continue to prove medium and long-term efficiencies and validate it with radiologists, clinicians and imaging researchers. Once we are able to do that effectively, the market will adjust accordingly, it always does.

The value seems to be in real-time results that can help in both safety and efficacy outcomes. Please expand further on this thought.

Real time results are everything if you really think about it. The main benefit is obviously how this technology facilitates the real time patient care capabilities of healthcare practitioners. Safety and efficacy instant results from treatments are critical when analyzing real time clinical research data. Imaging biomarkers can and will play a tremendous role in improving patient standard of care, and ultimately can make a real impact on disease outcomes.

by: Dan Sfera, also available at: https://medium.com/quibim/quibims-strategic-advisor-in%C3%A9s-perea-discusses-the-radical-disruptions-that-quantitative-imaging-54cef3a3e082

Imaging Biomarkers in Lymphoma

By: Angel Alberich-Bayarri

7 July, 2017

Imaging has a crucial role in Lymphoma management nowadays. The main applications are based on the evaluation of disease extension in staging and in treatment response evaluation. Recently, thanks to the technology development of PET-CT and CT scanners, it has shown also a high utility in the evaluation of extra-nodular involvement, the early relapse and the transformation from indolent Lymphoma to an aggressive phenotype [1].

Evidence sets PET-CT and standard CT+contrast as the main imaging modalities for staging and treatment response evaluation. The most suitable modality will depend mainly on the aggressiveness and the FDG avidity of the lesion. Therefore, either for Hodgkin’s Lymphoma (HL), aggressive subtypes of non-Hodgkin’s Lymphoma (NHL) or for extra-nodal involvement evaluation in PET-CT will be the way to go for an appropriate staging. However, in cases of non-FDG avidity, mainly in indolent lymphomas (T-cell lymphoma and subtypes of NHL like Chronic Lymphocytic Leukemia, Marginal Zone Lymphoma, Lymphoplasmacytic Lymphoma), contrast enhanced CT is the main modality. Regarding response evaluation, a similar distribution of lymphoma subtypes per modalities is arrange, with the difference in the Follicular Lymphoma (FL), where PET-CT is the most suitable technique for those FL with a high tumoral burden, whereas low tumoral burden FL should be studied by CT with contrast when studying response. Up to now, Magnetic Resonance Imaging (MRI) has still not shown enough evidence in the management of lymphoma patiens beyond Primary Brain Lymphoma. PET-MR has a promising future in Lymphoma evaluation, specially in the current need for low dose follow-up studies that could be done with this modality.

Imaging Applications in Lymphoma

Due to heterogeneities in FDG metabolic uptake in different Lymphoma subtypes, Deauville criteria were established to grade the avidity in comparison with mediastinum and liver. However, conventional PET-CT has limitations in the staging of nodular alterations, with the exception of FL, where PET-CT helps to increase the stage of Lymphoma by detecting additional disease in up to 29% of cases. Regarding response evaluation, PET-CT has been recently considered as the gold standard at end of treatment in FL. This is one of the main conclusions from GALLIUM study.

Despite the previous comments, we understand that the staging and response evaluation in PET-CT in patients under new treatments based on targeted therapies or immunotherapy can not be only based on SUVmax evaluations. New imaging biomarkers have been developed in order to evaluate complex clinical scenarios like indolent Lymphoma or reactive inflammatory changes at the end of treatment in patients that have responded to therapy.

In the following table, specific Imaging Biomarkers for different biological objectives are provided:

Objetive	Modality	Imaging Biomarker
Tumoral burden	PET-CT	Metabolic Tumor Volume (MTV)
Tumoral burden + Metabolic activity	PET-CT	Total Tumor Glycolisis (TTG)
Change in metabolic activity	PET-CT	Voxelwise Delta-SUV (ΔSUV)
Heterogeneity	CT & PET-CT	Textures

For Imaging Biomarkers implementation, we always follow the step-wise method that we developed and published and that was also considered in this European Society of Radiology guideline.

The first technical step of Imaging Biomarkers development workflow after an appropriate definition of the idea is the Images Acquisition. In PET-CT, European Association of Nuclear Medicine (EANM) guidelines should be followed, and centres should be ideally certified by EARL program.

Metabolic Tumor Volume (MTV)

The first Imaging Biomarker to be calculated is MTV. It is defined by consensus as those lesion voxels with a significant FDG uptake, that is >41% of SUVmax although different thresholds can be evaluated in practise. The typical units are cm^3. The analysis is performed semi-automatically by thresholding and manual correction.

Calculation of Metabolic Tumor Volume

Several studies have analysed MTV values in different types of lymphoma, in the following table from Schöder H. J Clin Onc 2016, a nice summary can be appreciated:

Schöder H, Moskowitz C. Metabolic Tumor Volume in Lymphoma: Hype or Hope? J
Clin Oncol. 2016 Sep 6. pii: JCO693747. [Epub ahead of print] PubMed PMID:
27601547.

The different thresholds used for MTV can be also appreciated (although 41% SUVmax is the one in the majority of them). Also, the wide range of MTV obtained show us the high heterogeneity of the disease and raises also the concern about treatment dose. Should we modulate the treatment given to patient by the MTV? or, on another way, does a patient with 600cm^3 of MTV have to receive the same treatment dose than a patient with 3000cm^3 ? Important research needs still to be done in this field.

Regarding MTV and Follicular Lymphoma, few studies have been performed. The most important one was a retrospective analysis from Meignan et al, where they calculated a MTV of 510cm^3 for 2-year Progression Free Survival (PFS). However, some controversy has arose mainly due to the fact that the inherent error in SUV measurements due to examination variability introduces a final MTV error in measurements around 20%, so the threshold should not be a single value, but a given range of MTV values that consider that error.

Total Tumor Glycolysis (TTG)

The TTG measurements is better applied for specific lesions rather than all lesion burden. Therefore, if we focus on specific lesions, the TTG combines information on the FDG avidity and the MTV of the lesion by the following equation:

TTG = MTV x SUVmean

Voxelwise delta-SUV

The structural and anatomic information contained in the CT examination within the PET-CT acquisition can be used for spatial registration of scans of the same patient corresponding to different timepoints (e.g. registration of end-of-treatment CT on baseline CT). The idea behind this is to create a parametric map of the longitudinal SUV changes in the patient, and for that the deformation field resulting from spatial registration is applied to the end-of-treatment PET in order to convert it to the baseline geometry. After this process, the follow-up examination can be superimposed to the baseline and therefore even substracted to calculate the SUV difference between timepoints.

delta-SUV pipeline

Textures analysis

The image regions can be also evaluated quantitatively by means of texture analysis. Texture analysis allows for the extraction of quantitative descriptors from voxel intensities relationships within an image or region. They are organised in first order (if directly extracted from histogram) or second order (if an additional step is required for their calculation). Texture analysis and specially heterogeneity biomarkers like the entropy and kurtosis have shown promising results in many different cancerous lesions, specially as a prognostic biomarker.

Texture analysis from lymphoma lesion in CT

In lymphoma, a recent manuscript from Ganeshan B. et al. has shown excellent results in providing complimentary information to the interim PET as a prognostic biomarker.

However, texture analysis techniques can also be applied to other type of images such as the PET component, being able to determine the metabolic heterogeneity (MH) of the lesions. In this regard, lesions with different regional FDG avidity are having a worse prognosis than lesions with a homogeneous FDG uptake.

Metabolic Heterogeneity in FDG uptake in lymphoma

As you have discovered in this post, there are still many Quantitative Imaging Biomarkers that can be extracted from conventional FDG PET-CT examinations, and which are showing important relationship with lymphoma progression, according to recent investigations. In QUIBIM we are a committed team dedicated to the implementation of these techniques in clinical practise, research and clinical trials. If you want to collaborate with us in this field do not hesitate to contact us and potentially upload a case through our QUIBIM Precision® platform. It will be the best way to start working together in this emerging field.