Chapter 3
AI to Advance Medical Imaging
Medical imaging is probably the most promising clinical application of AI. Whether it is diagnosing a cancer, detecting a fracture, or identifying neurological or thoracic conditions, AI can help quickly diagnose and assist doctors with required treatment options.
Visualization of Class Activation Mapping (CAM) results. AI-based model assessments of different COVID-19 cases provide doctors with insights into the algorithm’s decision.
It is estimated that there are about 40 million radiologist errors per annum due to either overworked radiologists or poor quality of imaging techniques [4]. AI algorithms help radiologists make diagnoses by recognizing subtle anatomical structures and deducing clinical meanings. AI also helps by processing and providing analyses of large volumes of images in a much shorter time.
The use of AI in diagnostic medical imaging is undergoing extensive evaluation. As of July 2022, 75% (391) of the devices approved by the FDA in the market were in radiology imaging alone [5].
Challenge
Exposure to radiation from computed tomography (CT) imaging is approximately 350 times that of a single X-ray dose and is associated with several risks, such as cancer. Medical researchers want to limit radiation exposure by using ultra-low-dose CT scans. However, this approach results in low-resolution images with high levels of noise, which make scans difficult for physicians to interpret.
CNNs trained on ultra-low-dose CT. (Image credit: Ritsumeikan University)
Solutions
A researcher, Ryohei Nakayama from Ritsumeikan University in Kyoto, Japan, used MATLAB to create a deep learning convolutional neural network (CNN) that reconstructs high-resolution images captured using ultra-low-dose CT scans.
- At first, the researcher used MATLAB to divide CT images into small local regions and pair low-dose and normal-dose regions to create an image dictionary. As the dictionary grew, the search time became untenable, so Nakayama explored use of a convolutional neural network (CNN), which produces results much faster despite the training time.
- Nakayama used MATLAB to evaluate about 128 different CNN variants, trying different input sizes and filters as well as various numbers of convolutional layers.
- To accelerate the training process, he trained in parallel on multiple NVIDIA® GeForce series GPUs using Parallel Computing Toolbox™.
- To monitor training progress, Nakayama plotted accuracy and loss using the monitoring visualization option in Deep Learning Toolbox™.
Results
The CNN-based system provides physicians with a comparable level of diagnostic information while reducing patient radiation exposure by as much as 95%.
Challenge
Small lumps or growths on the thyroid gland are usually benign, but a small percentage could be malignant. Physicians use ultrasonography to diagnose thyroid nodules, but the accuracy of the diagnosis depends on the experience of the radiologist. Radiologists assessing the same nodule can sometimes arrive at different diagnoses.
Solutions
A research team at Yonsei University and Severance Hospital in Seoul, South Korea, used MATLAB to design and train convolutional neural networks (CNNs) to identify malignant and benign thyroid nodules. The researchers validated the CNNs against data sets from multiple hospitals, packaged them with a user interface, and deployed them as a web application, all using MATLAB.
SERA web app developed in MATLAB. (Image credit: School of Mathematics and Computing, Yonsei University)
- Initially, they used Statistics and Machine Learning Toolbox to perform feature engineering and train multiple machine learning models, including support vector machine (SVM) and random forest classification.
- The team then started exploring CNNs with Deep Learning Toolbox and worked with 17 different pretrained networks in MATLAB, including AlexNet, SqueezeNet, ResNet, and Inception.
- The 17 different networks were trained on a data set of more than 14,000 images. For feature-based combination, they used the outputs of the final fully connected layer in each CNN to train an SVM or random forest classifier. The weighted average of the classification probability produced by each CNN was calculated to make the inference.
- The trained CNNs were made available in the hospitals that Yonsei University works with by creating a web app named SERA and deploying it with MATLAB Web App Server™.
Results
Diagnostic tests have shown that these CNNs perform as well as expert radiologists. The application is used by medical students as part of their training and by experienced radiologists who need an objective second opinion on diagnoses.
Challenge
During the early days of the COVID-19 pandemic, detecting the coronavirus disease was difficult, especially with the large number of cases rising in the entire world.
Solutions
Detection and diagnosis tools offer a valuable second opinion to doctors and assist them in screening process. A researcher from the University of Dayton Research Institute (UDRI) used MATLAB and Deep Learning Toolbox to develop an automated deep learning algorithm for COVID-19 detection using chest radiographs. He also visualized the class activation mapping (CAM) results for the various trained networks for different COVID-19 cases to help provide insights behind the algorithms’ decision to the doctors.
Classification results for normal and COVID X-ray scans.
References
[4] Bruno, Michael A., Eric A. Walker, and Hani H. Abujudeh. “Understanding and Confronting Our Mistakes: The Epidemiology of Error in Radiology and Strategies for Error Reduction.” Radiographics 35, no. 6 (2015): 1668–1676. https://doi.org/10.1148/rg.2015150023.
[5] Center for Devices and Radiological Health. “Artificial Intelligence and Machine Learning (AI/ML) Enabled Medical Devices.” U.S. Food and Drug Administration. FDA. Updated October 5, 2022. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices
