THE ADVENT OF ARTIFICIAL 

INTELLIGENCE IN CANCER 
DETECTION 
A SURVEY  
─ 

Shreyas Akella 
UWM ID 991360075 

 
 1 
 

Abstract 
 
Cancer is a dangerous disease with evidence based discussions, suggestive of 
long-term effects. The evidence is usually generated from exterior of an organism, 
and it usually comes from extensive medical quantitative analysis. A few examples 
covered here will range from skin cancer to colon cancer. A number of trends are 
noted, including a growing dependence on protein biomarkers and microarray 
data, a strong bias towards applications in prostate and breast cancer, and a heavy 
reliance on “older” technologies such artificial neural networks (ANNs) instead of 
more recently developed or more easily interpretable machine learning 
methodsThis paper aims to discuss the areas where Machine Learning(ML) and 
Neural Networks(NN) and applicable and what is the result of doing so. This paper 
will also discuss the individual progress treatments, and the difference between 
incorporation of AI and the usual techniques used. In assembling this review, I’ve 
surveyed many papers (referenced below), while drawing conclusions from a few.  

Introduction 
 
Cancer is a very degenerative disease, and the prediction of this disease has been 
quite challenging for decades, until machine learning methods have been applied 
to cancer detection. The fundamental goals of cancer prediction and prognosis are 
distinct from the goals of cancer detection and diagnosis. In cancer 
prediction/prognosis one is concerned with three predictive foci: the prediction of 
cancer susceptibility (i.e. risk assessment); the prediction of cancer recurrence and 
the prediction of cancer survivability. 
 
There are a lot of factors, when it comes to detecting cancer. Most of the details 
usually lie with the patient, and they include height, weight, history, allergies, 
treatments for said allergies, smoking behaviour (for lung cancer), age, location and 
environmental upbringing. Other issues that do affect prognosis, but are out of the 
range of a patient’s control are the location of the cancer, whether it has 
metastasized, whether it is degenerative, etc. Usually, these details require very 
 2 
 

specific molecular details, about the patient’s genetic make-up or the tumor 
present.  
 
With the advent of new techniques like genome sequencing, and with the 
improvement of other techniques like tissue arraying, medical imaging, etc, these 
molecular details have been very easy to obtain. 
Artificial neural networks (ANNs) are regression devices containing layers of 
computing nodes (crudely analogous to the mammalian biological neurons) with 
remarkable information processing characteristics. They can detect sways from 
linear regressions that are not explicitly formulated as inputs, making them capable 
of learning and adaptability. They possess high parallelism, robustness, 
generalization and noise tolerance, which make them capable of clustering, 
function approximation, forecasting and association, and performing massively 
parallel multifactorial analyses for modeling complex patterns, where there is little 
prior knowledge. 
 

Specifications 
 
In this section, I have reviewed three different types of cancer, namely, colon 
cancer, breast cancer and skin cancer. The papers I’ve reviewed talk extensively 
about employing ANN ‘s and CNN’s (convoluted neural networks), in a way that 
addresses the prognostic definitions. Each of these cancers have a different 
approach towards artificial intelligence.  

COLON CANCER  
 
In this scenario, Artificial neural networks (ANNs) have been employed. They are 
rigorously adaptive and help the database capacity tenfold. They possess 
non-linearity, noise insensitivity and high parallelism. This helps to gain a better 
understanding of how the cancer advances, as a pattern, and helps identify the 
nature and any potential risks evaluated.  
 
 3 
 

When an ANN is given a particular set of data,it recognizes patterns in the data, and 
builds an analogous system to predict the same outcomes for that particular data 
set . Training, which is a synonym to biological learning, is carried by a "teacher" 
program that loads in training cases from a database and adjusts the weights and 
thresholds value of the network to minimize the error between the real-world 
outputs and the network generated outputs for the training case inputs. 
 
The benefit of an ANN is that it can learn from a multitude of variable analyses. 
Although larger data sets mean a bigger deviation from the actual probabilistic 
output, the neural network usually provides a method to minimize errors. This 
helps in the expansion of local knowledge to a more global database.  
 
As of now, it is quite tough to predict the advancement of colon cancer and its 
variants on the chemotherapeutic release. It is also difficult to predict the death of 
said patient with colon cancer at a quite advanced stage (B or C). 
 
In the particular paper I’ve studied, an ANN was trained from a 5 year follow up 
data set that included 334 patients with colorectal cancer. This ANN was trained on 
these data sets and another follow up for 12 months. This was to validate around 
50 patients and to predict death within 9, 12, 15, 18 and 21 months. This ANN was 
90% accurate when it came to predicting deaths. It had an overall accuracy of 79% 
when compared with clinicopathological methods.  
 
In the other study, in the UK, an ANN was used to predict the survival rate of 
patients suffering from colon carcinoma. An FFNN with 2 hidden layers with 4 and 3 
hidden neurons, respectively, was selected. The ANN yielded a ROC area of 87.6%. 
At sensitivity to mortality of 95%, the specificity was 41%. The logistic regression 
yielded a ROC area of 82%, and sensitivity to mortality of 95% gave a specificity of 
only 27%.  
 

SKIN CANCER 
 
 4 
 

The first part of a successful diagnosis of a malignant lesion by a dermatologist is a 

visual examination of the suspicious skin area. A correct diagnosis is paramount, 
because of the similarities of some lesion types; moreover, the diagnostic accuracy 
correlates strongly with the professional experience of the physician. Without an 
experienced doctor, the chances of a correct diagnosis is only about 65%-75%. 
 
Since the identification was so important, application developers and health 
specialists came together to develop apps on cheap and fast mobile devices, to 
deploy the approximations of machine learning on a large scale, employed by a 
large number of people. However, there were quite a few challenges. A lot of errors 
crept in, and false positives began showing up quite often. Another issue was the 
loss of image quality, and overall, this paved the way to inaccuracy. 
 
In 2016, the Symposium for biomedical imaging explored the depths of machine 
learning, and devised a way to predict the prognosis further accurately, by 
CNN(convoluted neural networks.) This is a part of supervised learning. It has 2 
major components, the HIDDEN layer and the FULLY connected layer. 
Convolutional Neural Networks, it was found, could adapt more quickly and learn 
through previous data faster.  
 
CNNs require sufficient training data to succeed at what the problem is. Due to the 
actual error rate produced with other learning techniques, CNN was deployed as it 
could learn the significant changes and colour complexity of the multitude of raw 
pixel data and correlate it with the class labels. This enables for feature extraction.  
 
In the paper I’ve reviewed, Kawahara. et. al used a linear classifier to classify 10 
different skin lesions. Feature extraction was also performed using an AlexNet 
whose last fully connected layer was replaced with a convolutional layer. Using the 
dermofit Public image Library, he acquired 1300 images of 10 different skin lesions, 
and fed them to the neural network. The result was an 81% accuracy in the skin 
lesion prediction. This gave way to another part of the prognosis:detection, and 
survival. 
 
 5 
 

Esteva et al, provided 129,450 different images of 2032 different skin lesions. Out of 
these, 3374 images were obtained from a dermatoscopic device. The trained CNN 
was tested with test data that were fully biopsy-proofed and achieved an ROC AUC 
of .96 for carcinomas, an ROC AUC of .96 for melanomas, and an ROC AUC of .94 
for melanomas classified exclusively with dermatoscopic images. 
 

BREAST CANCER 
 
Breast cancer is a survivable cancer that, if caught early, is treatable with a very 
high success rate. Unfortunately, early detection is quite a challenging process, and 
with the exception of self diagnosis, there was no way to tell if a patient has 
advanced cancer. If detected early the chance of survival goes from a mere 56% to 
a very high 85%. 
 
In the paper I’ve reviewed on breast cancer and its classifications, the proposed way 
to detection was the rule-based classification, for the prediction of survival.  
 
The data set used 8 attributes. This included a record of 900 patients, with 876 
being female, and 24 patients being male. 
 
Naive Bayes (NB), Trees Random Forest (TRF), 1-Nearest Neighbor (1NN), AdaBoost 
(AD), Support Vector Machine (SVM), RBF Network (RBFN), and Multilayer 
Perceptron (MLP) machine learning techniques with 10-cross fold technique were 
used with the proposed model for the prediction of breast cancer survival. The 
performance of machine learning techniques were evaluated with accuracy, 
precision, sensitivity, specificity, and area under ROC curve. 
 
Out of these patients, 93, unfortunately died, and 807 patients were alive. In this 
study, Trees Random Forest (TRF) technique showed better results in comparison 
to other techniques. The accuracy, sensitivity and the area under the ROC curve of 
TRF are 96%, 96%, 93%, respectively. 
 
 6 
 

This study demonstrated that Trees Random Forest model (TRF) which is a 
rule-based classification model was the best model with the highest level of 
accuracy. Therefore, this model is recommended as a useful tool for breast cancer 
survival prediction as well as medical decision making. 
  

PROBLEMS WITH NEURAL NETWORKS IN CANCER DETECTION 

I. SKIN CANCER 
 
The first issue with skin lesions was that the considered formulations of the 
images only slightly different. This wasn’t the case just with the training data, 
but the used data and the output as well. This meant a reduced accuracy, 
and reduced reliability on such incremental methods.   
 
Another challenge was the updating of the image directory. The lesions that 
were used as a part of the project only included images of light coloured 
people. This produced errors for people of different colour. An updated 
global image registry means a representation of the world population.  

II. COLON CANCER  

 
Due to a practical point of view, only a limited amount of data that may be 
related to the outcome of interest can be collected, and these data are 
mostly based on studies in which a standard regression model was used, and 
therefore only factors that were significant in a regression models are 
collected in subsequent studies. 
 
All variables and outcomes are measured with errors. This is a definitive loss 
on accuracy.  
 
 7 
 

There exist data barriers beyond which mathematical models are unable to 
make predictions in biological systems. 
Regression models are superior to ANNs when drawing inferences and 
interpretations based on outputs. 

CONCLUSIONS 
The  incentive,  to  improve  cancer  detection  through  neural  networks  isPlausible, 
based  on  the  positive  neural  adjustments.  Systematic  optimization  of  each  neural 
network  that  differ  in  their  approach  to optimization contribute to highly predictive 
network.  The  definitive  high  level  of  classification  accuracy,  sensitivity,  and 
specificity  indicated  by  the  results  has  the  potential  to  dramatically  increase  in 
Clinical  management.  Specifically  the  technology  could provide accurate prognostic 
Value,  as  a  precursor  to  histologic  or  cytologic evaluation of biopsied tissue helping 
on a pathological level. 
  
Unfortunately,  it  is  difficult  and  in  many  cases,  impossible  to  compare  the 
performance  of  published  classification  results  since  many  authors  use  nonpublic 
datasets for training and/or testing.  
 
While  ANNs  still  predominate  it  is  evident  that  a  growing  variety  of  alternate 
machine learning strategies are being used and that they are being applied to many 
types  of  cancers  to  predict  at  least three different kinds of outcomes. It is also clear 
that  machine  learning  methods  generally  improve  the  performance  or  predictive 
accuracy  of  most  prognoses,  especially  when  compared  to  conventional  statistical 
or expert-based systems. 
 

REFERENCES 
1. Skin Cancer Classification Using Convolutional Neural Networks: Brinker et al 
2. Neural Networks and Other Machine Learning Methods in Cancer Research - 
Alfred Vellido, et al. - University of Byrom St. 
 8 
 

3. Efficient Cancer Detection Using Multiple Neural Networks - IEEE Health and 
science Journal. John Shell, et al. 
4. Applications of Machine Learning in Cancer Prediction and Prognosis - Joseph 
A. Cruz, David S. Wishart - University of Alberta Edmonton - John A Cruz et al. 
5. Machine Learning models in Breast cancer survival predictions - Medical 
Informatics Research Center, Montezari.M et al. 
6. Artificial neural networks for diagnosis and survival prediction in colon 
cancer - Farid Ahmed - Dept of oncology, Brody School of Medicine