Original article can be found here (source): Deep Learning on Medium
Deep Learning Models For Medical Image Analysis And Processing
For applications like segmentation and disease detection
Deep Learning is now everywhere. No matter what the field, what the problem, there is always that one ML enthusiast who will try to fit the problem statement in a quaint deep learning solution. While deep learning’s application in some fields might seem unnecessary right now, one field that just cannot do without it in today’s time is healthcare. That’s right. Deep Learning has the potential to transform the entire landscape of healthcare and has been used actively to detect diseases and classify image samples effectively. Over time, these applications are only expected to grow. Nay, explode!
One such area is medical imaging. Medical imaging consists of a set of techniques to create visual representations of the interior parts of the body such as organs or tissues for clinical purposes to monitor health, diagnose, and treat diseases and injuries.
So, what is deep learning in medical imaging all about? Here are some applications in which deep learning is making huge strikes.
Medical Image Segmentation
Medical Image Segmentation is the process of identifying organs or lesions from CT scans or MRI images and can deliver essential information about the shapes and volumes of these organs. Earlier, the process of automating this procedure was done using edge detection filters and mathematical methods. However, with the advent of AI, deep learning has become the dominant technique as they demonstrate considerable capabilities in image processing tasks.
How is deep learning helping in Medical Image Segmentation
Deep Learning for Medical Image Segmentation has been there for a long time. Over the years, hardware improvements have made it easier for hospitals all over the world to use it. Convolutional Neural Networks (CNNs) are used for this process. CNNs work like a typical feedforward neural network, however, they are much more equipped to deal with images as they employ a mix of techniques such as convolutions, max-pooling, etc. The general idea behind using CNNs is to use a 2D input image and apply 2D filters (done using a 2D CNN) on it. Another approach is to use Transfer Learning, where models are trained by using pre-trained state of the art models and freezing the last few layers for learning weights specific to the problem. The low-level features are generally borrowed from ImageNet. Another unique approach is to use 2.5D CNN, which can deal with some amount of spatial information. They can strike a good balance between performance and computational costs. 3D CNNs provide yet better performance and deal with richer spatial information than a 2.5 CNN.
- State of the art 2.5 CNNs are used in knee cartilage segmentation
- With the availability of 3D imaging and improvements in 3D hardware, 3D models are now being widely used to segment brain tumors of arbitrary size.
- Fully Convolutional Networks, where the last fully connected layer is replaced with a fully convolutional layer, is used in multi-organ segmentation.
- Cascaded Fully Convolutional Networks are used for liver segmentation.
One of the most well-known structures for medical image segmentation is U-Net. It is built using the architecture of Fully Convolutional Networks. A modification, called 3D U-net, is used for vascular boundary detection. However, the most famous derivative of U-Net is probably V-Net, which applied the convolutions in the contracting path of the network, both for extracting the features and reducing the resolution by selecting appropriate kernel size and stride. To get started on medical image segmentation, any of these can be used for your next project with Transfer Learning. Here’s a link to a short tutorial on Transfer Learning to get you started:
If you aren’t aware on how to build your own CNNs or FNNs, this GitHub repository can be of great help in taking the first step:
Detection of Skin Cancer, Tumors, Diabetic Retinopathy
Anyone who has done the basic Andrew Ng course will surely remember him using the example of classification of tumors into malignant and benign using basic machine learning methods. State-of-the-art deep learning models are much more advanced though and are being widely used in cancer detection.
How is deep learning helping in detection of cancer and diabetic retinopathy
Diabetes Mellitus (DM) is a metabolic disorder in which pancreases cannot produce proper insulin (Type-1) or the body tissues do not respond to the insulin properly (Type-2) which results in high blood sugar. Due to the lack of experts on the issue, manual detection of diabetic retinopathy is a difficult and time-consuming process. Automated detection of DR based on deep learning models has proven their better accuracy.
Deep Convolutional Neural networks are widely used for detection of DR. For cancer detection, SVMs can be used for classification combined with CNNs for feature extraction.
AlexNet,VGGNet, ConvNEt-1 and ConvNet-2 can be used to build your own cancer detection systems. To learn more about how R-CNNs can be used to build your own object detection systems, follow the link:
Detection of Alzheimer’s and Parkinson’s Disease
Parkinson’s disease (PD) is a neurological disorder associated with a progressive decline in motor precision and sensory-motor integration originating likely from a disorder of the basal ganglia. Alzheimer’s disease is a progressive disorder that causes brain cells to waste away (degenerate) and die. Alzheimer’s disease is the most common cause of dementia — a continuous decline in thinking, behavioral, and social skills that disrupts a person’s ability to function independently.
How is deep learning helping in the detection of Alzheimer’s and Parkinson’s disease
Deep Boltzmann Machine (DBM) can be used for feature extraction and detection of abnormalities from 3D scans of MRI and PET images. According to Wikipedia, a deep Boltzmann machine (DBM) is a type of binary pairwise Markov random field (undirected probabilistic graphical model) with multiple layers of hidden random variables. It is a network of symmetrically coupled stochastic binary units. It’s okay if all of this is going over your head right now, it is just important to understand that DBM is a powerful Deep Learning tool for the detection of Alzheimer’s and Parkinson’s disease.
- DBM ran on the ADNI dataset obtained up to 96% accuracy.
- CNN models employed for the detection of a healthy or an Alzheimer’s brain obtained around 98%-99% accuracy
- Sparse auto-encoders obtained up to 87% accuracy in Alzheimer’s detection
- CNN models using GoogleNet, AlexNet, etc. have achieved around 98% accuracy in the detection of Parkinson’s disease
Boltzmann Machines and Sparse Auto-encoders are advanced concepts in deep learning which take a certain level of knowledge to get started. However, for a brief overview of how they work, you can refer the following links
Deep learning is indispensable to the medical industry today. However, what it has achieved is just the tip of the iceberg. Deep learning can automate every nook and cranny of the healthcare industry, and by expanding on this sector, it could help make healthcare accessible and affordable to all. Sweet, right?