Using XAI in to detect breast cancer

Project Title

Using XAI in to detect breast cancer

Project Area of Specialization Artificial IntelligenceProject Summary

Breast cancer is one of the most common reasons for the deaths in women. It not only threatens their physical health but their mental health as well. In the recent development in Artificial Intelligence with the availability of public datasets, AI has been deeply blended with imaging modalities like mammography, ultrasound, MRI, etc. to help physicians in diagnosing a disease. However, breast cancer detection through Computer Vision is affected by the image quality and the lack of model interpretability also makes the existing models to have low confidence. Hence, the need to use Explainable Artificial Intelligence (XAI) to achieve insight into how the classification of cancer occurs. We are working on De Novo Convolutional Neural Network (CNN) Architectures, which are Deep Learning models trained from scratch. The models use the CBIS-DDSM dataset which consists of 2,620 publicly available images. The dataset is already equipped with relevant annotations and specified Regions of Interest. We fed these images into a CNN, which classified the images accordingly. We have split the dataset of mammographic images into 90-10 Train-Val sets.

Project Objectives

Traditional Machine Learning algorithms are as complex as they may seem. They need a great deal of domain expertise, human intervention, and are only capable of what they’re designed for; nothing more, nothing less. In traditional machine learning algorithms, most of the applied features must be identified by a domain expert to reduce the complexity of the data and make the patterns visible more clearly for the learning algorithms to work. The biggest advantage of deep learning algorithms is that they try to gradually learn high-level features from the data. This eliminates the need for domain expertise and core feature extraction. We are trying to build a robust CAD model using de novo technique. Since our problem statement is to detect cancer in breasts, we are working on getting good results on our own architecture. This research aims to build a proof of concept that will allow researchers from both the Artificial Intelligence and Medical fields to prove that an assistive model can help diagnose breast cancer through mammograms at the earliest possible stage from a scientific perspective. This will be achieved through the following objectives:

• Collect and analyse information about the general attitude of clinicians who specialise in breast cancer in the use of artificial intelligence to assist them in the detection of breast cancer.

• Research into the functionality within the hidden layers of existent CNN models and of our own on how they predict whether a mammogram contains radiologically detectable foci of DCIS or not. We will give particular attention to Resnet50, VGG16 networks to compare our model’s performance.

• Employ the use of XAI to generate a rational justification of how images are processed through the neural network’s different hidden layers. 

Project Implementation Method

A manifest file for images was downloaded from the source in .tcia format which were extracted in dicom format through NBIA Data Retriever. A DICOM image file is an outcome of the Digital Imaging and Communications in Medicine standard. Specifically, image files that are compliant with part 10 of the DICOM standard are generally referred to as “DICOM format files” or simply “DICOM files” and are represented as “.dcm". The images are then converted into png format. PNG (Portable Networks Graphics) supports greater color depth and because it is usually lossless compression, it enables better image quality. To get baseline results, we first augmented the dataset in two ways:

• Histogram Equalization. It is a digital image processing method to improve image contrast. It does so by constructively spreading out the most frequent intensity values, i.e. expanding out the image’s intensity range.

• Image Sharpening. Kernels, also known as masks and convolution matrix, are used in the process of image sharpening. The technique of sharpening, image detection and others, require that a kernel should be applied to the pixels of the image. This is why this method is also known as Convolution- the process of kernel being applied to the image.

Since our focus is binary classification, i.e. benign and malignant, using sigmoid activation function is the best choice; where the model will predict 0 if the output class is benign and predict 1 if the output class is malignant. Our initial CNN architectures with their results are as as follows:

• De Novo Architecture with Image Sharpening. 

• De Novo Architecture with Histogram Equalization. 

• Transfer Learning with Image Sharpening.

• Lastly, we will be using the python library for XAI.

Benefits of the Project

Breast cancer is the second leading cause of cancer death in women. (Only lung cancer kills more women each year.) The chance that a woman will die from breast cancer is about 1 in 39 (about 2.6%).Since 2007, breast cancer death rates have been steady in women younger than 50, but have continued to decrease in older women. From 2013 to 2018, the death rate went down by 1% per year.These decreases are believed to be the result of finding breast cancer earlier through screening and increased awareness, as well as better treatments.

For breast cancer diagnosis, Computer Aided Diagnosis (CAD) method has been devel- oped to increase the efficiency and effectiveness of breast screening. Conventional CAD systems are limited to manually extracting features to describe suspect structures in the breast. Currently researchers aim to develop fully automated end-to-end CADS systems.

It is possible to get a fully automated CAD system with the rise of deep learning techniques, in which the features are automatically discovered through supervised learning. Particularly, Convolutional Neural Networks (CNNs) are widely being used for developing CAD systems for detecting and classifying breast cancer. A. Shrestha et al. developed a new algorithm to describe deep learning. The models included InceptionV3, DenseNet121, ResNet50, and VGG16 models, for the classification process.

The sudden progress and wide scope of deep learning, and the resulting surge of attention and multi-billion dollar investment, has led to a virtuous cycle of improvements and investments in the entire field of machine learning. It is now one of the most popular areas of study world-wide . Healthcare providers generate and record massive volumes of data containing extremely significant signals and information, at a rate which exceeds what the "conventional" methods of analysis can process. Machine learning therefore quickly enters the picture, as it is one of the best ways to integrate, analyze and make predictions based on large, heterogeneous data sets (cf. health informatics).

Technical Details of Final Deliverable

We will be proposing the de novo CNN architecture to classify breast cancer as benign or malignant. The model will have different convolutional layers, pooling layer, dropout layer and finally classification layer. 

A Convolution layer which is the foundation layer for the Convolutional Neural Network. It has a set of kernels which are learnt throughout the training. filters parameter is an integer value which determines the number of filters that convolution layer will learn from. Additionally, it also tells the number of resultant filters in the convolution. Next, kernel_size determines the dimension of the kernel. The padding parameter can be either ’valid’ or ’same.’ When the parameter is set to "valid", the input volume isn’t zero-padded, and the spatial dimensions are allowed to shrink naturally by convolution. Lastly, activation_function is a basic function that converts its inputs into outputs with a specific range of values. The final value given out by a neuron is determined by this function. We have set this value to be relu which is Rectified Linear Unit Activation Function. If the relu function receives any negative input, it returns 0; however, if the function receives any positive value x, it returns that value. ReLu can be written as: f(x) = max(0, x)

After the convolution layer, there is a MaxPooling2D Layer. This layer downsamples the input along its spatial dimensions (height and width) by obtaining the maximum value for each channel of the input over an input window (of size determined by pool_size). strides are used to adjust the dimensions of each window. 

Dropout Layer can be used with convolutional layer after pooling layer. It is a hyperparameter which determines the likelihood of the layer’s outputs being dropped out, or, conversely, the probability of the layer’s outputs being preserved. Its value varied since it was tuned to improve the results of our models. Flatten Layer flattens the input and it does not bring an impact on the batch size, The regular deeply interconnected neural network layer is the Dense Layer. It is the most preferred and often used layer. 

Lastly, we are going to use a python library for eXplainable AI.

Our deliverables will include a research document explaining the work we have done in order to achieve our goal. 

Final Deliverable of the Project Software SystemCore Industry ITOther Industries Medical , Health Core Technology Artificial Intelligence(AI)Other TechnologiesSustainable Development Goals Good Health and Well-Being for PeopleRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	80000
GeForce GTX 1050 Ti GPU	Equipment	1	54000	54000
External SSD 256GB	Equipment	1	6000	6000
Hospital Visits	Miscellaneous	1	2000	2000
Stationary	Miscellaneous	1	1000	1000
Technical workshops	Miscellaneous	1	7000	7000
Training	Equipment	1	10000	10000