Artificial Intelligence: Deep Learning to Accurately Predict Breast Cancer Diagnosis- Featuring CNN
A new revolution of technology is here
--
3500 years ago, the first case of breast cancer ever was diagnosed. A description of a bulging breast with little to no cure at the time, the Ancient Egyptians were stumped. Recorded in the famous Edwin Smith Papyrus, an ancient record of signs and symptoms of the spinal column and spinal cord, there was very minimal understanding of the condition. [1] 3,500 years later, scientists have considered breast cancer to be treatable in the earliest stages, but advanced forms are dangerous and less optimistic.
With every problem, there is a solution: recent deep learning technology has been developed to predict breast cancer abnormalities accurately.
But wait. Why could this be useful?
In Radiology, it’s observed that diagnostic errors for breast cancer patients can range from 5% to as high as 28% annually [2]. With breast cancer being one of the leading causes of death for women worldwide, there is a deep urgency to reduce this number and facilitate effective diagnosis for patients who require urgent care.
What if we create a future that would make this number as close to 0 as possible? It just might be worth it…
A key to unlocking this potential lies in artificial intelligence’s deep neural networks. With the already immense transformation of the healthcare industry, creating a fully functioning program that can analyze mammogram scans to extract suspicious findings, this technology has the potential to revolutionize breast cancer research. If detection happens at earlier stages because of AI, all healthcare forms could have near-zero misdiagnosis rates.
As a quick overview and a helpful way to navigate throughout this article, feel free to go through as you please:
Understanding Breast Cancer
What is the process of developing breast cancer? What are things to look out for and what make a tumor cancerous?What is Deep Learning?
Learn about the general structure of neural networks and a bit of applied mathematicsDeep Dive: CNNs
What are they? How do they generalize images for a computer to recognize?Breast Cancer X CNN
How breast malignancies are determined with CNN and what are the revolutionsWrapping up
Public opinion on using artificial intelligence in healthcare and what to look out for in the near future
Understanding Breast Cancer
Typically starting in the inside the milk ducts or the milk-producing lobules of the breast, breast cancer occurs when cells grow uncontrollably, creating a mass tissue- commonly described as a tumor. Breast cancer spreads to lymph nodes or other organs in a process known as metastasis: typically, at stage IV, this type of cancer is fatal and very difficult to un-spread.
There are two simple classifications of breast tumors: benign (noncancerous) or malignant (cancerous). The difference is hard to distinguish when only subtle differences are present.
At first glance in the image above, it is difficult to tell what characteristics might alert healthcare professionals to notice a difference. The lesions- abnormal tissue areas inside or outside the body- seem pretty similar, but there are a few indicators radiologists look out for.
What do the professionals see?
Radiologists use kinetic maps- texture features created from kinetic parameter maps calculated based on MRI scans- to clearly distinguish HER2-positive and HER2-negative breast cancer: benign and malignancy.
Movement
- Movement on a Kinetic Map → Limits reliability because motion registers as a change between dynamic sequences
- This is a measure radiologists take into account to reduce the impact of misreading a scan
Patterns
- Grouped areas of calcifications, especially regarding marked background enhancement or suspicious findings on mammography
- Tight clusters with irregular shapes and refined appearance could indicate suspicion
Symmetry
- Commonly, breasts are symmetrical, but asymmetry is not abnormal and is commonly noncancerous; however, differing breast densities could be a risk factor for breast cancer
- Maximum Intensity Projection: Uptake = None, Minimal (<25%), Mild (26–50%), Moderate (51–75%), or marked (>75%)
A heart-touching story ❤
At the young age of 40, Stephanie Cosby had her first ever mammogram screening and was delivered with heartbreaking news. According to the results of her breast biopsy, she had a small lump within her breast tissue, indicated as early-stage breast cancer [3]. Considering that she was very healthy for her age, Cosby had an increased chance of beating the cancer.
She met with a genetic counselor to find the best treatment options for her specific type of cancer and was tested for the presence of common breast cancer genes: BRCA-1 and BRCA-2, which widened her options for treatment.
She decided to go through with the option of bilateral mastectomy with reconstruction; the surgeon used her belly fat to reshape her breasts, and she recovered in just eight weeks.
“I feel like I can tell women “don’t wait!” My outcome was the result of very early detection” said Stephanie Cosby.
At the time, her mother was diagnosed with metastatic breast cancer, and while dealing with her own health, she watched her mothers worsen. Sadly, after three weeks of Cosby’s surgery, she witnessed her mother’s death. She learned so much from her personal journey and speaks about how her life changed. She advocates for understanding breast cancer’s risks and the normality of a mammogram: if caught in early forms, breast cancer is curable.
“I’m not a breast cancer survivor. I’m not “surviving.” I’m LIVING! In honor of my mom, I now live every day to the max” said Stephanie Cosby.
Self-checks can be strong indicators
The USPSTF- the United States Preventive Services Task Force- recommends that women aged 50 - 74 get mammograms every two years because they are at most risk [4]. However, women could also self-examine in the comfort of their homes [5]. Here are the steps:
1] Touch: Using one hand and lying down with a pillow at the arch of the back allows for a quick self-check. Moving the hand around in little circles with different amounts of pressure and slowly moving throughout the breast area to feel all levels of breast density. Examine for new lumps, thickening, hardened knots, or any significant breast changes in a vertical up-and-down motion.
2] Look: By standing in front of a mirror and bending forward, examining for factors like:
- Shape -> Has there been recent distortion or change in size?
- Skin -> Any rashes, redness, dimpling, or fruit-peel textured appearance?
- Veins -> Are there any newly formed bulging or increased number of veins?
Imprinting standard check-in patterns for women in the high-risk category could be life-saving. Stressing once again: the absolute benefit of early diagnosis increases survival chances by an astonishing 90%.
What are the scientific terms for these patterns?
With the already complex nature of breast cancer, cancer is in different shapes, sizes, and specific locations for everyone. We precisely measure the growth patterns of a tumor by analyzing the morphology. There there are three main types: Focus, Non-Mass Enhancement (NME), and Mass.
So what are the differences? The left side of the image above displays foci: these cancer cells are less than 5 mm in size and typically don’t occupy space. They are cells that can be seen only microscopically and stand out from normal breast tissue because of stains and appearance. In the middle, we observe Non-Mass Enhancement cancer cells: patchy areas of enhancement that don’t conform to a space-occupying lesion. They have an abnormal look on an MRI and don’t conform to a specific shape. The last part (mass) is associated with a cancerous tumor. On the right is a mass: this includes proper borders and pushes out against the tissue; it has a defined shape and is clearly visible in the MRI image.
The bottom line: breast cancer is very complex with many different forms and ways to get treatment, there is no guarantee that radiologists could possible examine every detail to give a perfect analysis each time.
Where could we possibly go from there?
Embark: The Deep Learning Model
It’s mostly what it sounds like: a program model that is able to learn an input and produce a (mostly) accurate output. But that is the most surface level it can get; deep learning is so much more. In this case, it is a sense of hope. Hope to improve early detection to lower breast cancer’s mortality rate AND improve patient satisfaction with accurate diagnostic results.
First: what is deep learning?
In the most bare-bones form, deep learning is a subset of machine learning. Similar to the way our brain contains millions of interconnected neurons that work together to process information, deep learning categorizes data into neuronal organization called neural networks to make predictions based on multiple levels of abstraction (layers) to represent data. By processing nonlinear transformations, it creates a statistical model to iterate through processing layers until it reaches an acceptable level of accuracy [6].
The specifics of the way deep learning models can calculate their outputs require statistics, linear algebra, probability theory, and calculus. There is a wide variety of complex concepts, but let’s take a look at the fundamental ideas that neural networks use to calculate accurate outputs.
Linear Transformation: captures linear relationships between input features so the model can understand the foundation of more complex relationships.
- Formula: z=Wx+b
- z is the weighted sum of inputs x, where W is the weight matrix and b is the bias vector.
Activation Function: introduces non-linearity into the model, enabling it to learn complex patterns that are more relevant in capturing the complexity of real-world patterns and structures.
- Formula (Sigmoid): s(x) = 1/(1 + e^-x)
- Formula (ReLU): f(x) = max(0,x)
Loss Function: quantifies the difference between the predicted output of the model and the actual target values (accuracy measure).
- Formula (Mean Squared Error for Regression): L(y,y^)=1/2(y−y^)²
- Formula (Cross-Entropy for Classification): L(y,y^)=−∑iyilog(y^i)
Backpropagation: efficiently computes gradients in a neural network by applying the chain rule of calculus. Iteratively adjusts weights and biases based on the computed gradients and allows the network to learn and adapt to the patterns present in the training data.
- Formula (Chain Rule): (∂x)/(∂L)=[(∂z)/(∂L)][(∂x)/(∂z)]
To sum up: There are many different functions and formulas that could be for deep learning based on a given input, and they all work together to achieve the highest accuracy possible.
Deep Dive: Convolutional Neural Networks
There are many types of deep learning models, but they all have the same basic structure using layers and neurons. Since we now understand how they work, let’s focus on a specific type of deep learning algorithm: Convolutional Neural Networks (CNN).
QUICK: What is the image below a picture of?
Hopefully, you can recognize it is a doodle of a flower; typically, when we look at an image, our brain is able to quickly process what the image represents. But computers don’t have the same ability- they require complex algorithms and math concepts for even simple classification tasks. That is why CNNs are so powerful: they specialize in pattern recognition for computers.
Video by IBM Technology provides a great overview of how CNNs work:
Let’s walk through the specific steps of CNN’s feature extraction, classification, and distribution.
Local Receptive Fields: In a CNN, small regions of input layer neurons are connected to neurons in the hidden layers (local receptor fields), which are translated across an image to create a feature map from the input layer -> hidden layer neurons.
Weights and Biases: Each concealed neuron identifies a common feature, such as an edge or a blob, in various parts of the image. The model systematically traverses specific pixel arrays, comparing them to a filtered image to identify similarities. This makes the network tolerant to the translation of objects in an image.
Activation and Pooling: Uses activation functions to modify the output of each neuron, mapping positive values to the highest point and assigning zeros for negative outputs. The activation step’s output can be additionally modified through a pooling step, which condenses the output of small neuron regions into a single output, reducing the dimensionality of the featured map. This simplifies subsequent layers and decreases the number of parameters the model needs to learn [7].
How can CNN help Breast Cancer Detection?
Currently, trials are being conducted to determine whether deep learning can outpower health professionals or if it can be a helpful tool in the industry.
CNN’s ability to feature extract enables it to enhance and easily understand malignancy in breast masses.
Since cancer is determined by cancer cell growth, current measures for examining mitosis are done manually by a pathologist looking at multiple high-power fields on a glass slide under a microscope- this is very time-consuming and labor-intensive. However, the problem falls here: this process is a good predictor of tumor aggressiveness, so finding a way to make this process faster and just as practical is necessary.
To do just this, A study conducted by Haibo Wang- a Case Western Reserve University senior researcher- and his team [8] created a max-pooling-based convolution neural network for mitosis detection in breast imaging. A CNN model was necessary because computers cannot differentiate between the highly variable shape and appearance of mitoses. With a series of convolution and pooling layers, it was trained to pixelate the image and classify each pixel as either mitosis or non-mitosis to determine abnormal cell division.
This image displays the complex process of distinguishing between the highly varying shapes and sizes across mitotic nuclei within the same image.
The Steps:
1] Converting the images above from colorized images to blue-ratio images, Wang and his team were able to assign the individual pixels to a higher blue intensity- which is capable of highlighting nuclei regions.
2] Training: Their model was trained using the stochastic gradient descent to minimize the loss function:
3] Testing: They set their model to output a malignant screening if the probability threshold for miotic nuclei surpassed 0.58.
“An exponential function is applied to the log-likelihoods of each candidate nucleus belonging to the positive (mitosis) class…to calculate the probability that it is mitotic”. [8]
4] Features: Wang’s team handcrafted features to be classified into three categories: morphology, intensity, and texture. Morphological features are derived from the binary mask of the mitosis candidate to capture aspects of the mitosis shape. Intensity features encapsulate statistical characteristics of mitosis intensity, while texture features represent textural attributes of the mitosis region. Principal component analysis was applied to calculate the number of handcrafted features, preserving the most significant features contributing to 98% of the unlimited componential variations.
5] Overcome inaccuracy: Reduce bias
(1) Replaced overlapping nonmitotic nuclei with their clustered center
(2) Applied synthetic minority oversampling technique for miotic cells
(3) Threshold of 0.58
Using other performance measures and random forest trees for classification accuracy, their results are as follows. They tested several different models, but their best method in classifying tumors in the Scanner Aperio Data Set was with handcrafted features + CNN. It produced an 84% precision and 65% recall. Their approach created an F-measure- statistical notation to measure a test’s accuracy- of 0.7345, which compares very closely to other methodologies previously researched. Still, Wang’s provides a faster option that requires far less computing resources.
Though this number may not be the closest to 100%, this technology has so much potential. It saves tons of computing time, and pathologists screening for mitosis tests can utilize this technology to its fullest advantage. As research is conducted and time passes, it will only go up from here.
While this is just one example of how CNN can be utilized for breast cancer for a specific input, mitosis of cells, CNN can also detect many possibilities from mammogram or MRI imaging.
Outpower or Facilitate?
As we just saw, CNN can be a powerful tool in image processing, but an important question is: Can deep learning surpass radiologists’ accuracy in breast cancer diagnosis?
“While real-life clinical radiologists are essential and irreplaceable, a clinical radiologist with the data, insight, and accuracy of AI will increasingly be a formidable force in patient care,” said Dr. Katharine Halliday.
A study of more than 80,000 women with an average age of 54 were split into two groups: one with AI + Radiologist and one with Radiologist + Radiologist. The standard radiologists classified 203 cases of breast cancer, while the AI-assisted group found 244 cases of breast cancer- a 20% increase in detection rate over the standard radiologist + radiologist group. Both groups were observed to produce the same amount of false positives: 1.5% [9].
The group with AI had an overall reduced screen-reading workload by an impressive 44% compared to the conventional group. However, radiologists are still critically necessary to make accurate and timely diagnoses of breast cancer. AI will not outpower radiologists any time soon but can be used as a significant marker to facilitate faster and more precise diagnosis.
Ethics: The public opinion
While rapidly growing with a lot of research and funding, there is still skepticism regarding the complete dependence of artificial intelligence on this essential subject. There is worry that unnecessary technology interventions could cause more problems with people receiving misdiagnoses.
“Continued concerns have been raised regarding potential harm associated with unnecessary biopsies and surgeries that are triggered by imaging findings in patients who do not have breast cancer.” [10]
Since deep learning cannot perform at full capacity and still produces errors, leaving diagnosis to experienced radiologists is best for now. This is not to say that deep learning cannot assist in breast cancer detection: leveraging this technology as a double check or second opinion would allow for further investigations if the human reader and machine disagree on a diagnostic result.
The journey towards fully leveraging artificial intelligence in breast cancer diagnosis remains ongoing. Still, the path is promising: with the potential to improve accuracy, reduce misdiagnosis, and ultimately save lives.
Empowerment and Sensibility
Learning one and facilitating the other. While breast cancer has been studied for many decades, we are in for a long ride not only for the power of deep learning but also for education on early detection benefits.
Whether it is further research, learning applied math, or adding more hidden layers for image classification, deep learning will continue to heal and elevate humanity every single day.
Understanding the inner workings: we are part of a greater movement in how machines and humans can revolutionize healthcare and beyond. There are causes for concern, but by exercising caution and putting the lives of patients’ sensibility first, deep learning could facilitate further developments for hospitals in need and everything in between.
