Medical image reader powered by artificial intelligence
Date
2023-09Publisher
Brac UniversityAuthor
Palok, Tanvir AhmedAhmed, Symum
Anim, Golam Kibria
Ratul, Shahed Sharif Bhuiyan
Alam, Shahrear
Metadata
Show full item recordAbstract
Misdiagnosis in medical imaging is a critical concern, risking patients’ health due
to the pivotal role of radiologists’ accuracy in diagnostics. Current cross-checking
methods for radiologists’ decisions are limited, potentially leading to errors and
treatment delays. This study introduces a data processing technique and an advanced
prediction system for improving disease detection accuracy in medical images.
Our main goal is to contribute to healthcare by developing a system capable
of achieving human-level or higher accuracy in disease detection across diverse
medical image types. To achieve this, we utilize deep learning techniques, specifically
Convolutional Neural Networks (CNNs), and leverage Transfer Learning with
pre-trained models. Data processing plays a crucial role, given the importance of
image availability and quality. We apply image enhancement techniques such as
Histogram Equalization, Adaptive Histogram Equalization (AHE), and Contrast
Limited Adaptive Histogram Equalization (CLAHE) to enhance image quality and
augment a limited training dataset. The advanced ensemble approach significantly
enhances the overall accuracy and reduces individual model variance. Validation
of our approach using confusion matrices reveals that selective class-wise voting
achieves the highest accuracy at 95.27% on the testing dataset. Additionally, our
customized weighted voting approach achieves an accuracy of 94.07% on the test
set. These results emphasize the effectiveness of our ensemble techniques in improving
disease detection accuracy. Our ensemble techniques offer substantial accuracy
improvements, promising more accurate and reliable medical diagnoses