This article was originally published in issue 105 of Position magazine. It is authored by Dr. Petra Hemholz, Senior Lecturer Photogrammetry in Spatial Sciences at Curtin University, and co-authored by Prof. Gareth Baynam (WA Department of Health), Dr. Richard Palmer (Curtin University), Paula Fievez (FrontierSI), Dr. Lyn Schofield, Dylan Gration, Dr. Cathryn Poulton, and Yarlalu Thomas, all from the WA Department of Health.
A rare disease is defined as such when it affects fewer than 1 in 2,000 people.
With over 7,000 known rare diseases, while individually rare, cumulatively they are common. Worldwide, an estimated 6-8 per cent of people suffer with some form of rare disease. In Australia, 1.5 million live with a rare disease, as do 30 million people in Europe, and 25-30 million in the US. Thirty to fifty per cent of these individuals experience their first symptoms during childhood. Rare diseases are complex, chronic, and often multisystem involving disability, suffering, and death.
Thirty-five per cent of babies die before their first birthday and 30 per cent of children before their fifth. Thirty per cent of rare disease patients are referred to an average of six different specialists. Some wait up to 30 years before receiving an accurate diagnosis, and nearly half of all initial diagnoses are incorrect. The need to quickly and accurately diagnose rare diseases is thus a global health priority.
Undertaking a diagnosis usually involves first analysing a patient’s phenotype – how they appear, how they behave, how their physical and cognitive abilities compare to their peers, etc. Approximately one third of genetic and rare diseases manifest as atypical facial traits. However, recognising salient aspects of the facial phenotype requires expert knowledge about how distinct versus subtle facial characteristics appear, and the ability to identify overlapping facial traits of clinical significance.
To assist in describing and understanding these facial traits (and the human phenotype more generally), the Human Phenotype Ontology (HPO) was created. This is a structured set of well-defined terms to describe phenotypic abnormalities and a tool for annotating and analysing human disease. The HPO database is comprised of 11,000 terms (as of 2015) which are used in 250,000 disease annotations for over 10,000 rare and common diseases.
HPO terms are widely used in the rare disease community for developing computational tools for clinical differential diagnostics and for phenotype-driven analysis of genetic data. By agreeing upon such a library of terms, clinicians are better able to share their findings among their professional networks in a precise and unambiguous manner – leveraging each other’s knowledge and enhancing the ability of each clinician to deliver accurate and timely diagnoses.
When undertaking diagnostic inference, Australian clinicians are faced with several challenges during the collection and analysis of patient information. Such challenges include:
1. The need to precisely measure many different aspects of the face.
2. A paucity of documentation about facial anatomy and measurements.
3. Understanding the nature of atypical measurements for ethnicities other than Caucasians.
Currently, facial phenotyping for clinical diagnosis is performed by manually taking anthropometric measurements using simple instruments such as plastic rulers, tape measures, or callipers. However, manually taking facial measurements can be imprecise and sometimes traumatic, for young children especially. In addition, in some cases, identifying the presence of atypical phenotypic traits is based less on the precise measurement of features and their comparison with known facial norms, and more on the clinician’s personal knowledge and experience.
This element of subjectivity introduces a not-insignificant source of bias into the diagnostic process. Due to the length of time needed to take many facial measurements, clinicians also often concentrate on taking measurements of features that already appear somewhat unusual – a kind of confirmation bias.
If subsequent measurements are needed, the patient must be brought back into clinic so that the additional measurements can be made. By taking standard photographs of a patient’s face, further measurements can be made without needing the patient to be returned to clinic, but lighting and issues with pose together with the lack of precise information about the geometry and form of the facial surface limits the clinical utility of 2D images.
Cliniface is a platform for the visualisation and analysis of 3D facial images and enables the objective detection of standardised phenotypic descriptors (HPO terms) to support clinicians in providing timely and accurate diagnoses. It utilises well established spatial methods and approaches for this work.
Cliniface uses 3D facial data captured by specialist camera systems (e.g., the Vectra H1 camera). In general, such systems can generate 3D surface scans with submillimetre accuracy (depending on the capture technology) providing a detailed representation of the facial surface as a triangulated texture mapped mesh (≈200,000 data points).
Image capture itself often takes only as long as needed to take a regular 2D photograph. The cost of such specialist camera systems is decreasing and the technology itself is being miniaturised so that in the next few years, it should be possible to acquire highly detailed and accurate 3D scans from commodity hardware such as might be found in future generations of smartphones. Cliniface is agnostic to how a 3D image is captured and can import images stored in a variety of widely used file formats.
Cliniface annotates the 3D surface with standard anatomical facial landmarks which are used to automatically extract measurements of potential clinical significance (typically straight line distances between pairs of landmarks). The landmarks are detected semi-automatically and their precise placement can be adjusted by the user without involving the patient further.
After providing simple demographic information (sex, date of birth, ethnicity), Cliniface analyses the measurements against its database of matching norms and flags those that are atypical for the patient as well as the HPO terms they are associated with. The analysis can be exported to file as human readable structured text (XML or JSON formats) for processing as part of other third-party analysis tools.
As well as an objective analysis, Cliniface can accentuate aspects of the 3D facial surface via selectable visualisations to help the clinician more easily appreciate curvature and (a)symmetry through the medial plane which can offer further clues and assist in arriving at a provisional diagnosis – subsequently refined through clinical consensus and corroborated by genetic testing. Based on the definitive diagnosis, a treatment plan can be developed.
The long-term goals of Cliniface are to support clinicians with:
1. Developing new objective measurements and norms.
2. Developing models of normal face variation for specific cohorts (e.g., for ethnicities other than Caucasian).
3. Assessing facial change over time for treatment monitoring.
4. Undertaking more sophisticated forms of analysis by incorporating data and algorithms from other research entities.
Cliniface’s main user group is clinicians who use the software for disease diagnosis and treatment monitoring. However, Cliniface’s extensibility (realised by its plugin architecture) means it can be used to conduct research in the area of facial anthropometrics and to explore the utility of new state-of-the-art 3D image processing and analysis algorithms. Thus, Cliniface offers computer vision and image processing researchers a low-effort path to putting their work directly into the hands of medical practitioners – those who are best placed and most eager to benefit from applied advances in the field of 3D facial image processing and analysis.
The team has been able to establish international partnerships with China, Japan, USA, India, Czech Republic, Belgium, and the UK. Cliniface is installed (without needing elevated user privileges) as a standalone application and all analysis is performed locally; no images or patient data are uploaded elsewhere. This means that a user’s local ethical and organisational/jurisdictional conditions about data sharing are automatically respected helping to facilitate strong and transparent collaborations between the Cliniface team and its users, and the tool is free and open source.
Further information is available at www.cliniface.org.