Author

Anup Khanal

Document Type

Thesis

Date of Award

2024

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

KC Santosh

Abstract

The field of medical image analysis has benefited significantly from developments in deep learning (DL), especially in convolution neural networks (CNNs). Even after these advances, there remains a need for more robust and accurate methods for various tasks such as classification, segmentation, captioning, and object detection, particularly in medical diagnostics. Inspired by the principle of the wisdom of the crowd, feature fusion is a common process used, especially in tasks related to classification. This thesis introduces a novel attention-based feature fusion architecture designed to enhance classification performance by integrating the Convolution Block Attention Mechanism (CBAM) before fusion. This approach selectively refines informative features, improving representation and classification accuracy. Additionally, using the inter-annotator agreement to measure homogeneity between backbone networks during feature extraction is proposed to aid optimal backbone selection. The effectiveness of the proposed model is demonstrated through extensive experiments conducted on two different forms of classification tasks on two different modalities of datasets: Fractured limbs and ChestX-ray14. Results from the Fractured limbs dataset indicate demonstrable performance improvements, surpassing the baseline fusion method by 0.5-2.5% and achieving an accuracy of 97.5%, precision of 95.6%, recall of 95.9%, and F1-score of 96.1%. Similarly, in the ChestX-ray14 dataset, the proposed model showed notable improvements in AUC per label across various pathologies, achieving an average AUC of 0.81. This performance is competitive with current state-of-the-art models and represents a ∼ 5% increase over the standard baseline fusion method. In a nutshell, this work highlights the potential benefits of incorporating attention mechanisms, such as CBAM, to refine the feature maps before fusing them to enhance diagnostic accuracy, especially in the classification task. To the best of our knowledge, this is the first approach in the use application of attention mechanism and inter annotator agreement during the process of feature concatenation on features extracted from base networks.

Subject Categories

Computer Sciences

Keywords

Attention Mechanisms, CNNs, Feature Fusion, Inter-Annotator Agreement, Medical Image Classification

Number of Pages

67

Publisher

University of South Dakota

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.