Chapter

36

Artificial Intelligence
be either rewards or
INTRODUCTION Series of reinforcements. which can learns from
punishments. In supervised learning. a machine
AtlicalntcllencAl)reprevents the capacity of rmachines semi-supervised
the inputs from human interpreter. In
andotware lomimic theupnitive functions of hurman-like learning. fewer inputs from human interpreter are provided
leamngand problcm-olving, Therc havc been strides in this The results of machine learning must be
to the machine.
arca over thc last Sew decades which has brought artificial validated for accuracy against real worldgiving the 'ground
Itelligence into our homcs. Be it smartphones, smart truth for Al performance.
Icleviaon%, cll driven vchicles, weh carch, language processing.
0t VIal processing, wc arc swept by the wave of artificial Deep learning: Deep learning is a subset of machinelearning
intellicncc in our daily Iife. In this article, we will focus on and is a basis for Al tools. Deep learning means a machine
artilicial intellipcncc andits inpact on medical imaging. has multiple layers of algorithms that enable them to learn
and take decisions on its own.

HISTORY Artificial neural networks (ANNS): Artificial neural networks

are formed by multilayered algorithms. ANNs are composed
In 1950 Alan'Iuring first described the Turing test whereina of millions of nodes or units which are connected by the link.
computer passes thc test if a human interpreter cannot These links create a network that can be fed forwards (where
dllerentiate if the response is from computer or a human. connections are in one direction) or recurrent (where there is
Theternn 'Artificial Intclligence' was first coined in 1956 in
a summer workshop at Dartmouth College in Hanover by feedback in loops). In radiologic imaging, the ANNs must be
trained with data sets which are hand-labeled image data sets
Jobn MeCarthy, an American computer scientist. The
levelopment of a weh scarch enginc named PubMed' by used by the algorithm to match the ground truth. The problem
the National Libray of Medicine in 1960was an imporlant of 'overfitting arises when new data is added to the data set.
dipital resource that lead to acceleration of development of The solution to it is 'augmentation' wherein new variables
biomedicine. The period from 1970s to 2000s was termed are added to the data sets to enhance diagnostic performances.
Alwinler as the rescarch was slowed down due to lack of The more the layers of algorithms and the more the rounds of
lundnp Bul since the 2000s the tables were turned for AI. A training, the performance of the network is better.
soliware company, IBM developed a supercomputer named
Applications in Radiology
Walsou' which compcted with human participants to win a
Iclevision contest, "Jeopardy" in 2011. In 2017, Baker used Computer-aided detection (CAD) is the earliest application
the Watson computer toidentify RNA binding proteins which of assisted reporting. CAD is not true Al as it makes diagnoses
are altered in anyotrophic lateral sclerosis. for which it is specifically trained and follows arigid scheme of
Machine learning (ML): ML is pattern identification and recognition. True Al can do autonomous learning without explicit
programming at each step just like humans. CAD was first
analysis where the machinecan improve performance from introduced for the detection of breast cancers in mammography.
the provided data sets. The lcarning can be supervised. It has also been applied to colon and lung cancer
reinforced, unsuperviscd, or semisupervised. In unsupervised screening.
learning, the machine learns patterns in the input without any Radiomics: Radiomics is the extraction of quantitative
fecdback. In reinforced learning, the machine learns from a information from the radiologic image which a human eye
666
 Radiormics has reccntly bccn thc mt
rescarchin,Ain
perceive rmcdical ismaging The tumot
and proypss
rot 1of decide the aygrcssivcncs inforrnatin
caan
LCNElly helps in identifying thc
analysis cyc. Using decp Current use of Al in Radiiy
lexture
assesscd by thc nakcd
CXtrapolatcd to rncasure
he
cannot
dala can he
algorithms,the mctastascs.
Segmentation und volume anulysis: in27FDA
probability.chances of dislarnt
ing
slgnancy
LIent and can be
response
becormbincd with rnolcular data
algorithrns can Inpare
theerapy.The.autormatcd
zeedgeneevaluatethe Response Evaluation
Criteria in
and (RECIST) and WHO criteria for
Ircatrnent
Tumours
stcps of
radiomics arc:
alid Thebasic Lung nydules: Varius t #are 1ave a to
ponse.
Featur Statsta Can detect srnall lung layratesrs
Image extraction anaysi3
segmentation
mage
a0U'Ston
data
Additionalsteps
can be image registration, smoking hisory. farnily histry
modules that are modality
denoising and other
nd application-specific.
(ormatting, CT scoring of COVID 19 infection: bwz s te tz
guidelines are pandermic affecting the lung partyra sry y
optimisation: The dose optimization were developed based n the tret ílgivoh
Dase However,thereeis variation in
practíce from institute
andItechnicians in Autormated software Were develupei wzlzz ne y
Al can help radiologists tracking patient's
spcific.
levels.
. SCore from the CT images.
personalized patient protocol, radiation risk with
institute
o
choosingaj exposure and Bone age estimation: There have been i soiaz ct
providing cumulative can determine bone age based on the fusion of ysis
patientsusceptibility.
dose,

Siructured reporting:
Many international societies have accuracy as good as a radiologis Mary hospials 2v:
initiative by proposing adopted the use of the softwzre.
jointlystarted
a structured reporting

BA(G T1 (9y (11ym)

BA SO C61 y
CaA NA
D.
BA (TWJ 1134
Age: 1031 y

L
Scor mples of clinical uses of Al in radiology, (A) shows automated bone age estimation of achild. B) Al based CI seveny
ofOf braie
brain ío 9patient with lung involvement. (C)Automated sectoral map of CT coronary angiography. (D)Al basec segmeniaton
calculation of brain volume.
668 Tertbook of Radology fot Resdernts and Tecnnicions
to the natient
data /s not transferred
Challenges in Al earned from the use of
who owns the data.
Medicotegal uspecs:Thc biggest hindrance in the acceptance
ofAlin routine useis the medicolegal imnplications that arise The Future Path
after misdiagnosis. The developer has to claim the which has FDA-approved
Applelaunched series-7 watch detection. however. its
responsibility and bear the expenses tothe patients. Most of ECGmonitoring and
arrhythmia apn can
the current devclopers are not in a state to take medicolegal receive clearance.The
not
Oxygen monitoring did there have been cases.
where
responsibility. Currently. allthe scans have to be read by a your physician and its
radiologist, This adds to cost of health care. notify
survived myocardial infarction due to
patients have notification. The radiologist
Radiomics: Although a very promising area. there are various prompt detection and early training In the curriculum to
b:
shortcomings for its application to routine practice. The will have to include AT research and make AI more
the
variability in data if the image acquisition parameters are able to participate in If theradiologist has to
altered or the scanner is different. Most of the studies have a acceptable and economical to use. diagnosis made by the
the
take legal responsibility forlearn
Small sample size and it is verydifficult to oplimize image how the computer came
acquisition across different institutes. The population cohort computer. he she needs to
has to bc large and include various ethnicities. Data used in to the diagnosis.
one center or ethnic group are not equally reproducible on ?
other ethnic groups. Will Alreplace radiologists
big NO. But a radiologist
Data privacy and protection: The biggest concern for The answer to this question is apractice will always have an
research in AL is data privacy. The data is personal and is who embraces Al in his routine
doesnot. Al will be a helping
used for commercialwork. The data is stored at the hospitals edge over a radiologist who
the radiologists as wel
but the ownership of data is within the patient's rights. All hand for increasing the efliciency of
their imaging findings.
third-party useof data nceds an informed consent. The profit as increasing the accuracy in
 Chapter

37

Bone Age Estimation

show the dates of appearance ofthe primary 0ssification centres are iderntified by capital letters 2nd
l to378
O5sificationcentresinnthe various bones. Primary secondary o5sification centres are idertified by TueTas
3 4 5

shoulder. (A) 8th week of

Ta371. Ossfication of bonss of the Ife. fuse about the 45tn
léeB, 5th end 6th weeks of fetal
humerus 1 vears:
rife: (C,eth week of fetal lfe. (1) Head of
lesser tuberosity 5 vears:
Greater tuberosity 3years,
TWoepiphyses for acromion 15-18 years, fuse by 25th year: A

coracoid process 1year, fuses 15th years: (5) Root

2Mdoe cf 7
Inferior angle
f oresidprocess 17 years. fuses 25th years: (6) border of
fsaoue 14-20 years, fuses 22-25 years: (7) Medial E

sala 14-20 yeers. fuSes 22-25 years: (8) Sternal end of davicle
G
-20 years, fuses 25th year. 2

1 Fig. 3.3. Ossification of the wrist. carpus and hand. (A)Capitate

3
4months: (B) Hamate 4 months: (C) Triquetral 3years: (D) Lurate
5 45years: (E) Trapezium 6years: (F) Trapezod Byears (G) Scaphoc
6 years: (H) Pisiform 11 vears: ()) Metacarpals 10th fetal week
B (J) Proximal phalanges 11th fetal week: (K) Midcle phalanges
12th fetal week: (L) Distal phalanges 9th fetal week (M) Middle
phalanxof 5th digit 14th fetal week. (1) Lower end of radius
Fig. 37.2. O:sfraionof bones of the elbow. (A)8th week of fetal 1-2 years, fuses 20th year: (2) Lower end of ulna 5-8 years, fuse
te(Bjath vwEer of fetal life. (1) External epicondyle 10-12 years. 20th year, (3) Metacarpal heads 2.5 years. fuse 20th years
u9es 17-1 ars: (2) Internal epicondyle 5-8 years, fuses 17 (4) Base of proximal phalanges 2.5 years. fuses 20th years:
14 jeers1sllum 1-3 years. fuses 17-18 years: (4) Head of (5) Base of middle phalanges 3 years, fuses 18-20 years:
(atus 56uses 16-19 vears: (5) Trochlea 11th years, fuses (6) Base of distal phalanges 3 vears., fuses 18-20 vears: (7) Base
1on year. 9ranon 10-13 years, fuses 16-20 years. of 1st metacarpal 2.5 years, fuses 20th year.
669
 670 Textbook of Radiology for Residents and Technicians

2 A

3
A

Fig. 37.4. Ossification of bones of the hip. (A) 7th fetal week.
(1) 1st year, fuses 18-20 years; (2) 3-5 years, fuse 18-20 years; C
(3) 8-14 years, fuse 18-20 years.

5
B

pelvis. (A) Ilium 3rd fetal

1 Fig. 37.7. Ossification of bones of the Pubis 5th fetal month
month. (B) Ischium 4th fetal month. (C)
3 year. (1) Y-shaped
2
Note: Ischiopubic ramus fuses at 7th crest; (3)Anterior inferior
cartilage-2 or more centres; (2) lliac
tuberosity. All appear
iliac spine; (4) Pubic symphysis; (5) Schial
A B about puberty and fuse by 20-25 years.
Fig. 37.5. Ossification of bones of the knee. (A) 7th fetal
week; (B) 8th fetal week; (C) 5 years. (1) At birth, fuses 20th
B
year;(2) Tibial tubercle 5-10 years, fuses 20th year; (3) 4th year,
fuses 25th year; (4) At birth, fuses 20th year.

H 1
E
F
G
D C

Fig. 37.6. Ossification of bones of the foot. (A) Calcaneus 6th

fetal months; (B) Talus 6th fetal month; (C) Navicular 3-4 years;
(D) Cuboid at birth: (E) Lateral cuneiform 1 year; (F) Middle
cuneiform 3 years; (G) Medial cuneiform 3 years; (H) Metatarsal
shafts 8-9th fetal week; (lto K) Phalangeal shafts 10th week.
(1) Metatarsal epiphyses 3 years, fuse 17-20 years; (2) Proximal Fig. 37.8. Ossification of vetebrae. (A) Body 8th fetal week,
phalangeal base 3years, fuses 17-20 years: (3) Middle phalangeal (B) Neuralarch-one for each side 8th fetal week. (1)
Transverse
base 3 years, fuses 17-20 years; (4) Distal phalangeal base 5 years, process; (2) Spinous process; (3) Super surface of the body
fuses 17-20 years, (5) Posterior epiphysis of calcaneus 5th year, (4) Lower surface of the body. All appear at about 16
years and
fuses at puberty fuse by the 25th year.