Date of Award

Spring 2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Information Systems (PhDIS)

First Advisor

David Zeng

Second Advisor

Renae Spohn

Third Advisor

Julie Wulf Plimpton

Fourth Advisor

Mohammad Tafiqur Rahman

Abstract

Recent innovations in the field of deep learning and computational power, both in terms of hardware and software technology gave much-needed attention. Implementation of models by using the techniques has attracted a lot of researchers and is still one of the most widely used choices of cross-sectional researchers from medical as well as information system and science. The inherent characteristics of the data make contribution in determining the effectiveness of multi-stage learning in brain tumor classification by MRI medical images. Transfer learning models’ direct application is limited by variations in brain tumor MRI dataset’s image quality, resolution, contrast, noise, and anatomical diversity. Existing approaches often ignore how these data characteristics influence the model's ability to generalize and maintain high classification performance. In other words, CNN or any deep learning model demands sufficient pre-training work effectively in picture analysis. The use of transfer learning turned out to be an efficient approach in Learning models from scratch with high accuracy. Real images that have been used in the training of models during their learning on ultrasound do not seem to be as effective as in the non-medical images.

The dissertation discusses deep learning models for medical image classification and introduces an EfficientNetB2-GRU hybrid model with improved transfer learning methods. The ensemble method performed better than all the individual base classifiers trained on the whole MNIST dataset. A comparative analysis of three well-known CNN models—VGG-16, VGG-19, and ResNet50, which have all performed well in medical image classification and segmentation tasks. All the models have been trained on MNIST datasets to determine the importance of TL model effectiveness with the use of OCL and how it affects the data characteristics. The EfficientNetB2+GRU model is employed for the proposed hybrid model using the two-stage transfer learning method. In order to achieve the research goal, an initial timeline mapping background study was done first for the knowledge base of developing deep learning models and made acquainted with the recent advancement and issues of the field. The second activity entails analyzing and comparing another set of state-of-art deep learning models to ascertain the best of kind deep learning models and then proceeded to design an IT artifact that adhered to design science method. Lastly, an experiment was conducted for the test of the model trained sequentially for the identification of the Optimum Cutoff Layer (OCL) of the model, which was really used to compare the model of various image quantities of variable sizes to determine the effect on the efficiency of models.

The dissertation provides multiple contributions in the deep learning field and medical image analysis. In light of the new developments in model, the dissertation developed a new IT artifact to improve the deep learning models for the use of decreasing computationally intensive procedure involved in training of the small sample of the image data. From the theoretical perspective – a robust review of literature using the timeline method will contribute to enhancing the domain knowledge. From a methodological perspective – conceptualization of a unique IT artifact will encourage the model applicability for optimum training. From application perspective a detailed comparative analysis of medical image characteristics over the effectiveness of the deep learning model will provide informed decision-making capability for the benchmark utilization of the various image characteristics.

Share

COinS