|
|
Reconcillio™ - Uniting the Experts
Automated Generation of Consensual and Consistent Contours
The Task: Contouring
Contouring is the word for Segmenting (or
delineating or marking) a region of interest (lesion, organ,
and tissue type) in a medical image (CT, MR, Ultrasound etc.)
[Figure 1].
Contouring is a necessary task in many diagnostic, interventional,
treatment planning and post-treatment procedures. Manual contouring
is widely used for delineating lesions; the clinical expert
uses the computer mouse or a digital pen to mark the region
of interest. |
Figure 1. Brain MRI – Contouring is
highlighting the boundaries of lesions to mark them for measurements. |
The Small Problem: Contouring is Time-Consuming
Depending on the image type, the disease in focus, and the number
of slices to be contoured, the task of segmenting lesions and organs
can be quite time-consuming; several minutes per patient are very
common. The repetitive nature of the contouring task makes it a
very tedious job for highly skilled clinicians; this has multiple
disadvantages: first and foremost, patients have to wait to receive
the care they need, and the error probability increases due to the
tediousness of the task.
Automated and semi-automated software solutions are commercially
available for some clinical cases. Their performance, however, in
terms of accuracy and speed, varies considerably. Besides, for some
cases, there is simply no software assistance, such that clinicians
have to rely on tedious manual contouring.
The Bigger Problem: Subjectivity & Variability
In contouring medical images, there is generally no “gold
standard”, meaning that there is no 100% accurate
contour for a given lesion/organ of a given patient. Different
experts (radiologists, oncologists, pathologists etc.) contour
differently (Figure 2).
This well-known dilemma is called “inter-observer
variability” and constitutes a serious impediment
in medical imaging. It has been known since the 1950s [1].
The so-called inter-observer variability is sometimes so drastic
that “observers agree on only 50% of the total delineated
volume” [2] and in other cases the agreement can even
drop to 40% [3].
The same problem also applies to individual experts, who
can mark the same image differently when they observe it for
a second time. This is called "intra-observer
variability". Generally, the variability in
contouring originates in the subjective nature of the task
that, in spite of all the anatomical knowledge of the experts,
still causes alterations and inconsistencies to occur. |
Figure 2. Prostate Contouring – Different
experts may mark the same image differently. |
Interobserver Variability
- "The failure by the observer to measure or identify
a phenomenon accurately, which results in an error. Sources
for this may be due to the observer's missing an abnormality,
or to faulty technique resulting in incorrect test measurement,
or to misinterpretation of the data. Two varieties are inter-observer
variation (the amount observers vary from one another when
reporting on the same material) and intra-observer variation
(the amount one observer varies between observations when
reporting more than once on the same material)." [Source:
National
Library of Medicine] |
The Solution: Reconcillio
Looking at the small and big problems in contouring, one has to
establish the requirements for a solution; a software tool can overcome
these obstacles when it can offer the following features:
- High Speed – The software has to be
fast and extract the contour in a fraction of a second
- High Accuracy – Being fast does not
help if the result is not accurate, whereas accuracy isthe agreement
or overlap of the software result with the expectations of the
individual experts
- Trainability – The software has to be
trainable for each individual clinical expert such that he/she
can establish his/her own best contours; the training should customize
a contouring assistant for each individual expert.
- Consensus Building – a trainable software
that captures the individual contouring and editing preferences
of each doctor is the necessary step for building a consensus
contour; to remove the variability you have to capture
it first!
- Consistency Verification – a trainable
and consensus oriented software should also be capable of verifying
how consistent the user’s contours are.
Segasist Reconcillio is a revolutionary approach
to the auto-contouring of medical images. It uses the Segasist
Engine to create “Knowledge Maps” for each expert
user who works with the software. Over time, the Knowledge
Maps grow and converge toward a high agreement with the expert’s
expectations, in terms of where the contour should be. In
that point of convergence, Reconcillio becomes capable of
“Contouring Like You Would” for each user. Hence,
multiple contours can be extracted, reflecting the individual
differences; a consensus contour can be built and
offered to each individual user (Figure 3) to assess
his/her own sensitivity and specificity, as well as conformity
index.
Reconcillio does not stop at differences between doctors;
it can also assist the same doctor to become more consistent
in the way he/she contours. The verification by Reconcillio
can offer each expert a consistent contour (Figure
4) by employing the Knowledge Maps of that user.
[1] Gandevia B., Stradling P., Observer variation in the
tomographic diagnosis of tuberuculous cavitation. Tubercle
1957, 38:113-22
[2] Inter-observer variability of clinical target volume
delineation for bladder cancer using CT and cone beam CT,
F. Foroudi, A. Haworth, A. Pangehel, J. Wong, P. Roxby, G.
Duchesne, S. Williams and K.H. Tai, Journal of Medical Imaging
and Radiation Oncology 53 (2009) 100–106
[3] Inter-observer comparison of target delineation for
MRI-assisted cervica cancer brachytherapy: Application of
the GYN GEC-ESTRO recommendations, Johannes C.A. Dimopoulos,
Veronique De Vos, Daniel Berger, Primoz Petric, Isabelle Dumas,
Christian Kirisits, Carey B. Shenfield, Christine Haie-Meder,
Richard Pötter, Radiotherapy and Oncology 91 (2009) 166–172 |
Figure 3. Reconcillio dealing with Inter-Observer
Variability – The user contours a region (red). Reconcillio
extracts 4 contours based on Knowledge Maps of 4 users who
use the software (cyan) to build a consensus contour (green).
Figure 4. Reconcillio dealing with Intra-Observer
Variability – The user contours a region (red). Reconcillio
extracts another contour based on the Knowledge Maps of the
same user (cyan) to build a consistent contour (green). |
|
