Dissertation Title:  Automating the medical image classification process through deep learning

Authors: Hafsa LACI 

• Institution: University of Tirana,  Faculty of Economics, Department of Statistics and Applied Informatics

• Field of study:  Information Systems Economy

• Publication date: 11.06.2026

• The dissertation is published in Albanian.

© Copyright: Hafsa LACI 

Published by the University of Tirana
Based on legal acts, regulations and policies of the UT

👉 Click here to view the full dissertation (PDF)

 

Abstract

Deep learning models have played an increasingly important role in the field of medicine, enabling the automation of many processes that until now have been performed manually by doctors. Among these processes, the classification of medical images is of particular importance, as it helps to limit the subjectivity that accompanies their manual interpretation. However, the integration of these models into real medical contexts remains an open challenge. Most existing studies evaluate the performance of models under idealized experimental conditions, which do not reflect the real complexity of medical data. Furthermore, they abstract from important issues related to data privacy, the interpretability of model decisions, the costs of the necessary computational resources, as well as the environmental impact of the processes implemented.

This dissertation treats the classification process as a multidimensional problem, which goes beyond simple optimization of model performance. Its goal is to conceive, build and evaluate a structured pipeline for automating the medical image classification process in real clinical conditions, ensuring reproducibility, reliability and practical applicability.

The proposed USafePCOS-XNet pipeline includes the collection and structuring of medical data, their de-identification in accordance with HIPAA and GDPR legal standards, as well as the training, evaluation and interpretation of the results of a deep learning model supported by the EfficientNetB0 network. The application context is the classification of polycystic ovary syndrome (PCOS) using ultrasound images collected from a private clinic. The model is evaluated through a rigorous experimental protocol, which includes data splitting at the patient level, 5-fold cross-validation, and the use of appropriate metrics for problems with imbalance between classes (PCOS and NON_PCOS).

To increase the reliability of the results, the Grad-CAM technique is integrated, which aims to interpret the model's decision-making. In parallel, sustainability aspects, such as process duration, energy consumption and carbon footprint, are evaluated. This study proposes a comprehensive methodological approach for automating medical image classification, intended to be implemented in practice in a reliable and responsible manner.

Field:

Information Systems in Economics

Keywords:

Deep learning, medical image classification, PCOS, de-identification, interpretability (XAI), robustness

JEL code: C45, I11, I18

 

 

Abstract

Deep learning models have had a major impact in the field of medicine, enabling the automation of many processes that have traditionally been performed manually by medical experts. Among these processes, medical image classification is of particular importance, as it contributes to reducing the subjectivity associated with their manual interpretation. However, the integration of such models into real-world clinical settings remains an open challenge. The majority of existing studies evaluate model performance under idealized experimental conditions, which fail to reflect the real-world complexity of medical data. Moreover, they often abstract away from critical issues related to data privacy, the interpretability of model decisions, the computational resources required, and the environmental impact of the applied procedures.

This dissertation considers the classification process as a multidimensional problem, rather than a task limited to optimizing model performance. Its main objective is to design, develop, and rigorously evaluate a structured pipeline for the automation of medical image classification under real clinical conditions, while ensuring reproducibility, reliability, and practical applicability.

The proposed USafePCOS-XNet pipeline consists of a sequence of well-defined phases, including the collection and structuring of medical data, their de-identification in compliance with HIPAA and GDPR legal standards, as well as the training and evaluation of a deep learning model based on the EfficientNetB0 backbone, followed by the interpretation of the model outputs. The application context is the classification of Polycystic Ovary Syndrome (PCOS) using ultrasound images collected from a private clinic. The model is evaluated through a rigorous experimental protocol that includes patient-level data splitting, 5-fold cross-validation, and the use of evaluation metrics appropriate for classification problems with imbalance between classes (PCOS and NON_PCOS).

To enhance the reliability of the results, the Grad-CAM technique is integrated with the aim of interpreting the model's decision-making process. In parallel, sustainability aspects are evaluated, including process execution time, energy consumption, and the associated carbon footprint. This study proposes a comprehensive methodological approach for the automation of medical image classification, intended to be implemented in practice in a reliable and responsible manner.

Field of study:

Information Systems in Economics

Keywords:

Deep learning, medical image classification, PCOS, de-identification, interpretability (XAI), sustainability

JEL code: C45, I11, I18