Our recognition approach based on Bayesian Recognition Procedure (BPR) with one independent
criterion. The training sample designed by our own algorithm includes two object states (Lung Cancer and
Other diagnosis). That training sample calculated based on MRI image analysis with using Hounsfield Units
(HU).
Bayesian Recognition Procedure represented by the expression:
P( x0 , x1 ,..., xn f i) P( f i)
P( f i x0 , x1 ,..., xn ) , i {C , O} (1)
P( x0 , x1 ,..., xn )
For the case with one independent criterion, this formula (1) reduces to the form:
P ( x1 f i ) P ( f i )
P ( f i x1 ) , i {C , O} (2)
P ( x1 )
Here in the formula (2) we have the following: C – lung cancer; O – other diagnosis; P( x1 | f i) –
probability of occurrence of an event x1 for the case of distribution of random values of a class with lung
cancer or a class with other diagnoses; P ( f i ) – the likelihood of lung cancer or other diagnoses; P( x1 ) –
probability of occurrence of an event in the entire sample.
Note that, all the probabilities in formula (2) replaced by frequency distributions obtained based on
the training sample that we have. When a new research object arrives with an event x1 at the input of the
BPR, each of the two classes yields an answer according to the formula (2) and these two values compared
with each other. If the answer for a class with lung cancer is greater than the class response with other
diagnoses, then this research object belongs to the class of lung cancer and vice versa. When comparing two
answers according to formula (2), we see that the denominator reduced and only the numerator remains. The
remainder of formula (2) is further reduced.
Since we are dealing with one independent criterion, P( x1 | f С) is estimated by the frequency
m
jС
, where m is the number of values of j in the training sample for the class with lung cancer; m
mС jС С
- the number of analyzes with lung cancer in the training sample. The sum of all m is equaled to m .
jС С
m
jO
The frequency P( x1 | f O) is equaled . The sum of all m is equaled to m . Frequency
mO jO O
mС mO
P( f C ) . Frequency P( f O) . If we substitute all these values in
(mC mO ) (mC mO )
formula (2) for both classes, we get the following two formulas:
m mC m m mO m
jC jC jO jO
, (3)
mC (mC mO ) mC mO mO (mC mO ) mC mO
If we begin to compare the right and left values in formula (3), then we find that our BPR is
constructed by comparing only two values m and m .
jC jO
Now the main thing is to get the distribution m and m . We do not discuss the details of
jC jO
obtaining these discrete distributions, but just give one of them (Figure 1)
Figure 1
By comparing the values of the columns opposite to each other for the class with lung cancer and for
the class with other diagnoses, we obtained recognition probability for the lung cancer class - 59.39% and
for other diagnosis class - 58.84%.
In the near future, we hope to optimize the work of BPR and lay out a conditionally free version of
the utility, which will assess the state of the objects based on the MRI training sample.