Title : Computer-aided detection and classification of ischemic strokes from MRI images in LMIC– A retrospective cohort study
Abstract:
Ischemic stroke is a leading cause of morbidity and mortality globally. The high interfacing ratio of patients to professionals in Pakistan’s population is a great hindrance to access to care. This study intends to develop a computer-aided detection and classification system that works on MRI images to help with the time economy of professionals. It is also expected to aid with the accuracy and early detection of Ischemic Strokes. The evaluation of the developed system will be documented by calculating standard performance metrics when deploying the algorithm in a test set with cases that were not used in the training and development of the algorithm. This will be a retrospective crossectional study, which intends to analyze the agreement between the prediction of the algorithm and actual ischemic stroke identification in the test set, as well as a detailed case-by-case analysis of points of disagreement. The study also intends to provide estimates of the time economy of the algorithm. This is a retrospective cross-sectional cohort study design that will work on ten years of MRI scans of AKU patients.
Audience Take Away Notes:
- Although a few algorithms have been developed for the detection of Ischemic Stroke with accuracies of Random Forest Learning algorithm working with 68% sensitivity and Convolution Neural Network based algorithm with 85% sensitivity.
- It is stated that work can be done on improving the performance of such algorithms.
- There have been algorithms designed to classify into ischemic and hemorrhagic. These have been developed in fully developed countries and trained on data sets pertaining to specific populations.
- n machine learning, the NFL theorem states that within certain constraints, over the space of all possible problems, every optimization technique will perform as well as every other one on average. This entails that in improving the performance of an algorithm an important aspect would be training the algorithm on a dataset closer to the population it is being designed for.
- Our approach is working on a dataset from the population it is being designed for which should provide greater performance metrics. There is limited to no work documented on the fully automatic detection of types of ischemic stroke based on cutting-edge machine learning techniques.
- This study intends to document the basic performance metric of a preliminary algorithm designed for the same purpose.
