ASSIGNMENT 1 619(Artificial Intelligence)

Cancer, neurology and cardiology are the disease areas where AI tools are commonly used. This report primary focus on the concept of using AI applications in the case of breast cancer. “Breast Cancer” refers to the malicious tumor resulted from the unpredictable growth of cells that are generated in the breast tissue (Akay,2009). Breast cancers occur as a wide range of tumours that vary in appearance and biology(Budzik, Patera et al. 2019). Some examples of breast cancers include invasive ductal carcinoma invasive lobular and metaplastic carcinoma.The chief difference between malignant cancers and benign tumours is that tumour cells are confined to one mass while cells in malignant tumours will tend to spread outside the tumour mass and invade tissues in the body requiring more drastic treatments than surgery.

Each breast cancer has sub types and the sizes can vary significantly, this means that diagnosis is not straight forward requiring very experience specialists to make diagnosis but this takes time. Hence accuracy of diagnosis is paramount. Neural networks can be trained with data sets to build models to classify data very rapidly and can be tested to determine their accuracy and this could be a useful tool to speed up diagnosis. Microarray data is gathered by the selection of probes on tumor samples (Gavaert et al.,2006). Collection of both types of data is sequential.

An Artificial Neural Network takes radiologic data as input and result of biopsy as an output. ANN are the algorithms whose functionality and structure depends upon structure and behavior of biological neural network. Decision making, cognition and pattern recognition are the main computational problems that are focused by artificial neural network.

Problem Assessment

Breasts cancer can be both in men’s and women’s but women’s re more on the stake for being be effected by it.FNA cytology and core biopsy have been in use in Australia for many years for the treatment of visible breast lesions. Svnte Orell can be credited with fostering breast FNA cytology in Australia since the 1970s, and Orell and his associates’ Manual and atlas of Fine Needle Aspiration Cytology11 marks a landmark in FNA cytology.

Current national Accreditation guidelines for breast screen services12 state that: at least 75% of cancers are diagnosed without the need for a diagnostic excisional biopsy. FNA cytology specimens identified as identified as inadequate/ insufficient are below 25%. False negative rates for FNA cytology procedures are below 6%. Core biopsy specimens recorded. Breast lump is the most common breast-illness presentation. The cytology of fine needle aspiration is divided into five groups based on. On cytology, nine on histology were benign, while eight were malignant. We have modelled the breast cancer data and explore and validate those models. Therefore we have made a model of our data below with the help of FNA.

FNA technique is the fastest and least invasive approach. The aim of the project is to make diagnosis using fine needle aspiration quicker and more precise by developing an AI model capable of classifying cancer as malignant or benign based on the 9 characteristics values derived from the FNA biopsy study. Those characteristics are:

• Clump thickness
• Regular nuclei
• Mitoses
• Bland chromatin
• Cell-shape uniformity
• Marginal adhesion
• Bare nuclei
• Cell-size uniformity
• Single epithelial cell-size

As benign and malignant tumours are entirely different to each other so we may fin obstacles to differentiate them but it is very important to find the cause before going toward the treatment so, we need to differentiate them well first. It can be done by observing with the help of biopsy like CT-scan, MRI or FNA etc. but we are using here FNA. With the help of this tool we can observe the affected area. Sometimes we don’t get the perfect and accurate data needed to study the proper disease that might be another challenge while we are dealing with it. So for correct diagnosis would allow all 9 features to be modelled. AI techniques are extremely successful in classifying, based on various data characterstics and patterns. Many appropriate AI techniques are used in literature to model this data, including support vector machine (SVM) classifiers, supervised compact hyper spheres(SCHSs) (Tingting et.al.,2008) and fuzzy logic classify the data:

1. ANN (Artificial Neural Network)
2. KNN
3. Logistic Regression
4. Random Forest

ANN– the ANN classifier used supervised learning to evaluate the Wisconsin breast cancer dataset. The batch rate used during the apprenticeship was 1 and the number of epochs was set to 100 as no changes were made by increasing this however the training time increased significantly.

Each of the tumour features, the development ANN has 9 inputs one and a single output 1 each malignant, or 0 for benign. The ANN has 3 layers and the numbers of neurons in each hidden layer were varied during optimization from 16,8 and 1 to 12,6,1 afterwards.

Various combinations of relu, elu, sigmoid and tanh were used as activation functions:

• relu, relu, relu
• elu, elu, elu
• relu,relu, sigmoid
• elu, elu, sigmoid
• tanh, tanh, tanh

The neural sequential network has been tried with specific neuron numbers and activation mechanism for the 3 layers see results in table 1:

Table 1 ANN results:

The number of neurons in the layers varied attempting combinations of (24, 8, 1), (16,8,1) and (12,6,1). Differing the number of neurons in the layers made less than 1 percent difference for the results error, but (16, 8, 1) performed most consistently.

Activation roles best achieved were tanh-tanh-tanh and relu-relu-sigmoid. Tanh-Tanh-Tanh produced the best pattern. While reu-relu-sigmoid provided results that were marginally more liable. Through this may be because of the loss function, binary cross entropy with sigmoid performance better.

The accuracy of the ANN models created from each workout session varied considerably. Table 1 reveals that malignant rates range from 1.25% to 11.25% across different models. Model accuracy variance means several models need to be run to get a good one and this issues, take long training times, is becoming a problem. Fortunately, these problems can be overcome quickly by preserving a successful model so it can be used to automatically identify performance, avoiding more training steps.

K Nearest Neighbour (KNN)

The K-nearest neighbor algorithm is a machine learning algorithm. It is easy to implement and performs very complex tasks. It is a slow because it has no specialist training process. Then, it uses all the data for training when classifying a new data point.

Results in Table 2 show that K-nearest neighbour (KNN) model performed very well in accurately classifying the breast cancer data.

Table 2 KNN results

Model	% Accuracy	% Error
	Combined	Benign	Malignant	Combined	Benign	Malignant
KNN	95.61	95.95	95.00	4.39	4.05	5.00

KNN obtained overall accuracy of 95.61 percent with 4.05 percent benign error and 5 percent malignant error rivaling ANN.

KNN training took less than 3 seconds to complete very quickly and tests from each trained model were consistent at 100 percent. The results provided more reliable results than ANN due to the accuracy of the tests on average KNN models.

Logic Regression

It is a classification algorithm for machine learning, which is used to estimate categorical dependent variables likelihood. The dependent variable in logistic regression is a binary variable containing data encoded as either 1 (Yes) or 0 (No).

Table 3 summarizes the testing results for the logic regression classifier model.

Table 3 logic regression results

Model	% Accuracy	% Error
	Combined	Benign	Malignant	Combined	Benign	Malignant
Logic Regression	95.18	95.96	93.75	4.82	4.05	6.25

With a malignant error rate of 6.25%, the logic regression classifier achieved reproducible overall accuracy of 95.18 percent.

The model of logistic regression was easy to train and produced reproducible results. The feed data analysis shows that some of the variables are not fully independent, and the precision of the logistic regression may be enhanced by the number of variables input.

Random forest

Result in Table 4 show that the random forest classifier performed well with 10 trees achieving overall accuracy of 94.74 percent and 7.5 percent malignant error. Reducing the number of trees below 10 or growing them above 14 was detrimental to benign accuracy while malignant accuracy stayed constant between 6 and 19 trees at 92.5 percent.

Table 4 Effect of number of trees on Random Forrest classifier:

Model	No.	% Accuracy	% Error
	Trees	Combined	Benign	Malignant	Combined	Benign	Malignant
Random Forrest	6	94.30	95.27	92.50	5.70	4.73	7.50
Random Forrest	9	94.30	95.27	92.50	5.70	4.73	7.50
Random Forrest	10	94.74	95.95	92.50	5.26	4.05	7.50
Random Forrest	11	94.74	95.95	92.50	5.26	4.05	7.50
Random Forrest	12	94.74	95.95	92.50	5.26	4.05	7.50
Random Forrest	13	94.74	95.95	92.50	5.26	4.05	7.50
Random Forrest	14	94.74	95.95	92.50	5.26	4.05	7.50
Random Forrest	16	94.30	95.27	92.50	5.70	4.73	7.50
Random Forrest	18	94.30	95.27	92.50	5.70	4.73	7.50
Random Forrest	19	94.30	95.27	92.50	5.70	4.73	7.50

Like KNN and logic regression the random forest classifier was very fast to train and produced 100 % consistent results if the number of trees was kept constant.

Compared to other models, the random forest had the largest malignant error of 7.5 percent. This was also significantly higher than the 4.05 percent benign error, which can be a consequence for the model over suitable benign data.

Comparison of models

In table 5, results of the optimal test results are compared for each model.

Table 5 Model performance comparison results:

Model	% Accuracy	% Error
	Combined	Benign	Malignant	Combined	Benign	Malignant
ANN	96.93	95.95	98.75	3.07	4.05	1.25
KNN	95.61	95.95	95.00	4.39	4.05	5.00
Logic Regression	95.18	95.96	93.75	4.82	4.05	6.25
Random Forrest	94.74	95.95	92.50	5.26	4.05	7.50

From the results in Table 5 the order of the classifier accuracy form best to worst is: ANN, KNN, logic regression, and random forest.

The difference between the 3 models overall accuracy differs by 2 percent, but because there are more benign tumours than malignant ones, the [percentage error in the malignant tests gives a clear image of the different models results interestingly; the neutral error was the same as 4.05 percent for the 4 models.

ANN with just 1.25 percent malignant error performs far better than the other version. But the performance of the ANN model varies considerably, taking several training runs to get a successful model, whereas each time the other models yield reliably accurate models.

Component accuracy is limited to the size and consistency of the comparatively small original data set. It was also difficult to assess the consistency of the test results and could greatly affect the accuracy of the models. No model can be 100 percent, but any diagnostic mistake can potentially be devastating, leading to death or unnecessary surgery.

Summary

• Order of best to worst model performance: ANN, KNN, logical regression random forest.
• Results were interpreted by evaluating the models percent acc error overall and for both benign and malignant data on the test data collection. The ANN sequential neural network performed best for malignant cancer with a total accuracy of 96.93 percent and the lowest overall error. Interestingly, the neutral error was the same as 4.05 percent for the 4 models.
• KNN was similar to ANN with an overall precision of 95.61 percent and a error of 4.05 percent for benign and 5.0 percent for malignant tumours. KNNs training time was slightly faster than ANN took less than 3 seconds and the model was reproducible at 100 percent.
• While the ANN provided the best overall accuracy, the number of malignant cancer error models generated in each training session ranged from 1.25% to 11.25%. The ANN was also the slowest to train with the other models taking less training in under 3 seconds, taking about a minute. This may have been a issue an older computers and sluggish network connexions but modern computers are equipped with a decent GPU can handle the task in seconds.
• Learning time and reproducibility issues can be solved with the ANN by inserting code to save the best pattern.
• Although 96 percent accuracy is incredible, the effects of missed diagnosis range from needless surgery to death are tragic. Reducing the probability of missing diagnosis by adding a reliability score to the outcomes based on how well it suits the models can be done without 100 percent accuracy. To order to validate the diagnosis, unclear tests are then flagged for expert assessment or further study.