HealthManagement, Volume 16 - Issue 1, 2016
Where are we Now in the Algorithm Revolution?
The past forty years has been a period of unprecedented and
sustained advances in radiology with consecutive innovations expanding our
specialty’s reach and its diagnostic and therapeutic prowess. Technology has
been the midwife of our gains as more and more incisive capabilities have come
within our purview. But now we must reckon with the realisation that not every
‘improvement’ will be placed on our parochial agenda. Some may indeed threaten
our primacy even as they effectively improve patient care. The genie has come
out of the bottle, so to speak, eager to be deployed insistently and
decisively. One instance in which the threshold has been trodden, making the
challenge imminent and the disruption clearly evident, is the autonomous
application of the computer for the definitive diagnosis of conditions
affecting the brain and its coverings.
The integration of computers into our practice over the past
twenty-five years or so has been a felicitous development. They have been
incorporated into cross-sectional imaging devices, have made possible the
telecommunication of pictures as well as words, and have supported the voice
generation of reports. The virtues of these accretions to our clinical acumen
and our delivery of expertise are undeniable. More controversial has been the
conjunction of histologic references with pictographic patterns as evaluated by
computers in mammographic analysis. It is this application to which we have
affixed the term computer-assisted diagnosis, or CAD.
It is ironic that the first widespread linkage of morphology
to cell type with computers has taken place in the context of the diagnosis of
breast malignancy. The mammographic image is characterised by a spectrum of
shades of white, black and grey, making it difficult to distinguish abnormality
from normal, especially in dense breasts. Faint calcifications are also
discriminating, but here too some distinctions are not clear cut. Computer
assistance, its adherents maintain, helps bridge the gap between two realms of
spatial display— macroscopic patterns and microscopic cellular identity. The
accuracy of CAD remains a subject for continuing discussion. The problem it
addresses is unique—as providing a pathway for future refinements it is not a
dead end but rather a cul-de-sac. We must look elsewhere to evaluate
computerisation’s potential for furthering its integration into imaging
interpretation.
And that may now be happening. Recent augmentations in
computerisation power in the assessment of big data have focused
computer-directed analysis in novel ways. the claim being made now of CT and MR
evaluations of intra and extra axial lesions is that an unaided computer
investigation can make a diagnostic determination, not merely assist in one. in
the breast, computers are meant to link macroscopy with microscopy. In the
brain it is morphology alone for what they can now be tasked.
The brain is a rigidly circumscribed, symmetrical organ with
clearly delineated parenchymal conformations and intervening and surrounding
liquid spaces housed in an unyielding radiopaque shell. It is ideal for the
recognition of expansile, constrictive and eccentric abnormalities by
experienced interpreters. But now sophisticated computer algorithms, informed
by comprehensive databases of the brain, generated in various conditions and at
various ages, potentially offer a substitute means of pattern comprehension at
least equal to the interpretation of a radiologist.
That is the prospect. What to date is the evidence? A first
report from Japan in 2005, published in Radiology, assessed computer
‘assistance’ for the diagnosis of intracranial aneurysms by MR (Hirai et al.
2005). Only saccular or fusiform aneurysms alone were assessed. CAD was judged
better than radiologist interpretation but about equal with that of
neuroradiologists who did not avail themselves of the computer program. In 2010
a small series of patients with either intracranial, subdural, or epidural
blood collections revealed equal results by computer alone and by the
evaluation of a neurosurgeon (Liao et al. 2010). A retrospective review of a
computer algorithm to detect midline shift appeared in another article in 2010
(Xiao et al.). In 53 patients, the results had a sensitivity of 94% and a
specificity of 100%.
A more recent report regarding computer detection of stroke
lesions at CT showed that CAD proved useful for diagnosis of both haemorrhagic
and ischaemic strokes, and better for the detection of haemorrhagic lesions
(Gillebert et al. 2014). The authors focused on old atrophic brains. Most
recently, in 2015, computer diagnosis was assessed for tissue characterisation
of brain tumours by MR (Arakeri and Reddy 2015). This sophisticated program
considered shape, texture, wavelet and boundary characteristics. The
computerised interpretations equalled that of a neuroradiologist and exceeded
the evaluations of two less-experienced radiologists (Arakeri and Reddy 2015).
These reports together reflect the increasing capability of
computer determination. We emphasise here the thrust of these studies was for
computer determination not assistance. Clearly they betoken a compelling
alternative to conventional interpretation by qualified specialists. The payoff
could be great for those who can demonstrate that the technique could be made
available as a stand-alone exercise.
Furthermore, public perception may also play a role in the
ultimate allocation of proprietorship of imaging studies in which computer
determination will compete with diagnoses rendered by humans. The musings of
opinion makers in the general population are often influential in ultimately
directing both the choice of studies and the choice of caregivers responsible
for the studies so chosen. A seemingly gratuitous comment in a recent OP-ED
column in the New York Times by the noted geopolitical pundit Thomas L.
Friedman is germane (Friedman 2015). In an essay about the global agenda facing
a new president, he stated: "Robots are milking cows and IBM's Watson
computer can beat you at Jeopardy! [an American quiz show] and your doctor at
radiology" (Friedman 2015). Or to paraphrase, it will beat your
radiologist at diagnosis. So public audiences have been brought into the issue
by this comment. Will they soon insist on a computer-determined report as a
standard the ‘fallible’ radiologist may not be able to meet? That sounds
perverse perhaps, but once the matter becomes a topic for lay discussion it
cannot be ignored.
Moreover, it is likely that existing computer-determined
algorithms will improve. A recent announcement by IBM about Watson indicates
the company’s interest in applying it to imaging (IBM 2015). So where will that
situate radiology? For many CT and MR examinations of the brain, computer determination
will be situated initially within radiology's domain. But once it is realised
that the computer is doing the diagnostic work and the radiologist is now the
manager of the device, and not the interpreting clinician, other physicians
might seek to take the business away from us. In the United States, jurisdictional
boundaries demarcating specialists’ responsibilities are permeable. Neurologists
and neurosurgeons could soon realise that the radiologist’s interpretation may
then become superfluous for routine cross-sectional imaging analysis of the
brain. As long as procedure content and volume are directly related to income,
they will attract interest from those physicians who regard themselves as
conversant if not expert with the technique. For example, interventional
neuroradiology, once the province of radiologists, primarily has now become
populated in recent training programme classes by neurologists and
neurosurgeons, who consider themselves at least as capable as radiologists to
meet the subspecialty’s challenges.
Hence we must acknowledge and so confront the great changes
impinging upon us by the strident march of technology, no longer in step with
us, but quite possibly ahead of us, determining the path clinical diagnosis
will pursue. Will we stand aside or follow, or find another way to demonstrate
enduring value?
Key Points
- Computers have benefited radiology over the last 25 years,
including cross-sectional imaging, telecommunication and reporting.
- Computer-assisted diagnosis has been more controversial as
to benefits and accuracy.
- Computers can now provide analysis of brain morphology equal
to a neuroradiologist.
- Radiologists need to face the possibility of computer determined diagnosis, and patients may yet prefer it.
References:
Arakeri MP, Reddy G (2015) Computer-aided diagnosis system for tissue characterization of brain tumor on magnetic resonance images. Signal, Image and Video Processing; 9(2): 409-25.
Friedman TL (2015) Are you sure you want the job? New York Times, 21 October. [Accessed: 1 December 2015] Available from nytimes.com/2015/10/21/opinion/are-you-sure-you-want-the-job. html?_r=0
Gillebert CR, Humphreys GW, Mantini D (2014) Automated delineation of stroke lesions using brain CT images. NeuroImage Clin, 4: 540-8.
Hirai T, Korogi Y, Arimura H et al. (2005) Intracranial aneurysms at MR angiography: effect of computer-aided diagnosis on radiologists’ decision performance. Radiology, 237(2): 605-10.
IBM (2015) Watson to gain ability to “see” with planned $1b acquisition of Merge Healthcare [press release] 6 August. [Accessed: 1 December 2015] Available from www-03.ibm.com/press/us/en/pressrelease/47435.wss
Liao CC, Xiao F, Wong JM et al. (2010) Computer-aided
diagnosis of intracranial hematoma with brain deformation on computed
tomography. Comput Med Imaging Graph, 34(7): 563-71.
Xiao F, Liao CC, Huang KC et al. (2010) Automated assessment of midline shift in head injury patients. Clin Neurol Neurosurg, 112(9): 785-90.