Convolutional Neural Networks: How AI can save lives
What pops into your mind when you hear the words artificial intelligence? Many of us instantly think of some groundbreaking apps like chatGPT- a chatbot that answers complex questions in a surprisingly human way, or Dream- an app that can create incredible paintings from a user-entered prompt. Some of us film buffs would probably think of the Terminator (T-800) and how SkyNet will be our downfall! Now, even though these are very great examples of AI and its potential uses, what if I told you there’s a whole side of AI that we rarely get to see? The point I’m trying to make here is that a lot of us understand the uses of AI from an entertainment and command-response-based aspect, but we may not be aware of its emerging presence in other fields, Medical Imaging and Computer Vision in particular. This is where Convolutional Neural Networks come in.
Convolutional Neural Networks
A Convolutional Neural Network (CNN) is a deep learning algorithm that can take input images, assign importance to various objects in each image and be able to differentiate them from one another (1). The architecture of a CNN is similar to the connectivity pattern of Neurons in the human brain and was inspired by the structure of the Visual Cortex.
Why is this important?
While I won’t go into the processes of the deep-learning algorithm in depth, the main thing to take away from this is that the algorithm can extract objects from an image and assign significance to the objects. If you feed many images to the algorithm, it will be able to learn these features/characteristics. This process of passing more images to the algorithm so it can learn and distinguish features better is known as training. With enough training, a CNN will be able to recognize an image and accurately point out all the key features.
This technology has already had a profound impact on tasks where image recognition is vital, like biometric facial recognition and self-driving cars. It’s also beginning to make strides with medical imaging as I’ll discuss how in the next section.
Detecting Skin Cancer
Skin Cancer is caused by the rapid growth of abnormal cells in the epidermis, the outermost layer of skin. Such growth is caused by unrepaired DNA damage that triggers mutations that lead to malignant tumors. An estimated 1 out of 5 Americans will develop skin cancer by age 70. Even though it may not be as lethal as other forms of cancer, Melanoma, the deadliest form of skin cancer, is affecting more people at an astounding pace as it has been reported as the 6th most common of all cancer cases (2). The number of new cases of skin melanoma is 22.2 per 100,000 men and women per year.
Early detection of skin cancer is especially crucial because it can be easily treated/removed before progressing to more advanced stages. Currently, various forms of skin cancer are diagnosed visually. The process begins with an initial clinical screening and is potentially followed by dermoscopic analysis, a biopsy, and histopathological classification. A manual inspection by a dermatologist has a success rate of 65–80% and can be improved to 75–84% with dermatoscopic images (3).
With the assistance of machine learning, automated cancer classifications can help dermatologists by freeing up their time and providing fast and affordable access to lifesaving diagnoses, even from one’s own home, through the use of mobile devices.
Case Study
In recent years, Convolutional Neural Networks have been trained to distinguish between benign and cancerous skin lesions in various case studies. In one study, a CNN was trained using a dataset of 4,867 clinical images obtained from 1,842 patients diagnosed with skin tumors between 2003 and 2016 (4). The images consisted of various diagnoses and contained both malignant and benign conditions.
The overall goal of the study was to determine whether deep-learning technology can accurately classify types of skin cancer given a relatively small dataset of images.
The CNN’s performance was tested against 13 board-certified dermatologists and nine trainees. The overall classification accuracy of the deep learning algorithm was 76.5%, with the ability to classify malignant at 96.3% and benign at 89.5%. The CNN's overall ability to distinguish between benign and malignant tumors at 92.4% was even higher than the board-certified dermatologists at 85.3%.
It’s clear that Convolutional Neural Networks have the capability to classify skin cancer as accurately as the most qualified dermatologists in the world. Automating detection through the use of CNNs can save time for dermatologists and patients as well as avoid the costs of a traditional screening. In the near future, deep neural networks will be able to run on your smartphone, potentially providing low-cost universal access to vital diagnostic care.
This is just a microcosm of the use of machine learning in medical imaging, researchers have successfully used CNNs for many medical image understanding applications like the detection of tumors and their classification into benign and malignant, detection of optical coherence tomography images, detection of colon cancer, blood cancer, anomalies of the heart, breast, chest, and eye (5).
For more info about Convolutional neural networks and how they operate, check the following article: https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53
Resources
1. https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-networks-the-eli5-way-3bd2b1164a53
2. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7445643/#B16
3.https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6231861/
4.https://academic.oup.com/bjd/article/180/2/373/6601593
5.https://link.springer.com/article/10.1007/s12065-020-00540-3
