Leveraging power-of-two pixel value enhancement for enhanced feature contrast in medical imaging

다국어 초록 (Multilingual Abstract)

Medical imaging stands at the forefront of contemporary healthcare, playing a critical role in early disease detection, clinical diagnosis, and therapeutic monitoring. Despite rapid advances in deep learning--based image analysis, the preprocessing of high-intensity pixel regions, an essential aspect of identifying diagnostically relevant structures, remains a persistent challenge. Conventional methods such as histogram equalization, adaptive contrast enhancement, and basic morphological operations frequently fall short in effectively highlighting subtle pathological features, particularly within noisy or heterogeneous datasets. This inadequacy often results in diminished edge clarity and compromised model performance, thereby limiting the potential of automated systems to assist clinicians in real-world environments.

In this thesis, we introduce and rigorously evaluate a novel pixel-level preprocessing technique termed Power of Two Pixel (PTP) Value Enhancement, specifically designed to selectively amplify high-intensity regions within medical images. The PTP method leverages a nonlinear exponential transformation to intensify pixels above a defined threshold, thereby enhancing edge visibility and improving the contrast of critical anatomical or pathological boundaries. To provide an empirical basis for the approach, we formulate a dedicated Edge Sharpness Measurement (ESM) metric that quantitatively captures improvements in edge definition before and after enhancement.

The proposed enhancement strategy is systematically integrated with state-of-the-art convolutional neural network architectures, including ResNet50, MobileNetV2, and ConvMixer. It is evaluated across four publicly available and clinically diverse medical image datasets: diabetic retinopathy (APTOS 2019), skin lesion analysis (ISIC 2019), microscopic parasite species detection, and multi-class eye disease classification. Extensive experiments are conducted to benchmark the PTP method against standard preprocessing techniques and baseline models using accuracy, loss, precision, recall, F1 score, and area under the ROC curve (AUC) as evaluation metrics.

Results from both quantitative and qualitative analyses indicate that PTP Value Enhancement consistently leads to significant improvements in model performance across all tested datasets. In particular, classification accuracy increases up to 99.88\% on the parasite species dataset, accompanied by corresponding enhancements in precision and recall. The method also demonstrates superior generalization and robustness in the presence of noise and inter-dataset variability, as evidenced by higher ESM values and more stable training convergence. Furthermore, ablation studies on pixel intensity intervals confirm the effectiveness of targeting the upper quantile of intensity values, while comparative assessments establish the lightweight and architecture-agnostic nature of the proposed preprocessing framework.

This thesis contributes a robust and practical solution for pixel-level image enhancement, with demonstrated benefits for medical image analysis tasks that rely on edge detection and structural feature extraction. The findings underscore the potential for PTP Value Enhancement to serve as a plug-and-play preprocessing step in various medical imaging pipelines, thereby facilitating more accurate and interpretable diagnostic support systems. Future research directions include extending the methodology to segmentation and multimodal fusion tasks, as well as validating its applicability in larger-scale and heterogeneous clinical datasets.

목차 (Table of Contents)

• 1 Introduction 1
• 1.1 Motivation 1
• 1.2 Objectives of the Thesis 3
• 1.3 Organization and Contribution 4
• 2 Related Works 7
• 2.1 Advancements in Medical Image Analysis using Deep Learning 7
• 2.1.1 Innovations in Diabetic Retinopathy Detection 7
• 2.1.2 Classification Techniques for Diabetic Eye Diseases and Cataracts 10
• 2.1.3 Deep Learning in Parasite Microscopic Image Analysis 12
• 2.2 Pixel-Based Image Preprocessing in Deep Learning 14
• 3 Proposed Methodology 17
• 3.1 Detailed Methodology Breakdown 18
• 3.1.1 Data Collection 19
• 3.1.2 Power-of-Two Pixel Value Enhancement for High-Intensity Pixels 19
• 3.1.3 Model Training 22
• 3.1.4 Selection of Best Epoch 24
• 3.1.5 Evaluation on Test Set 24
• 3.2 Advantages of the proposed method 25
• 3.2.1 Enhanced Edge Detection 26
• 3.2.2 Applicability to Various Domains 27
• 3.2.3 Improved Image Quality 28
• 3.2.4 Efficient Feature Extraction 28
• 3.2.5 Robustness in Diverse Conditions 28
• 3.2.6 Transfer Learning Capability 29
• 4 Experimental Results and Discussions 30
• 4.1 Experiments and Results 30
• 4.2 Datasets 31
• 4.3 APTOS 2019 Blindness Detection Dataset 32
• 4.3.1 Parasite Species Dataset 33
• 4.3.2 Skin Diseases Dataset 34
• 4.3.3 Eye Diseases Dataset 34
• 4.4 Experimental Environment 35
• 4.5 Baseline Models 36
• 4.6 Evaluation Metrics 37
• 4.7 Edge Sharpness Measurement (ESM) Evaluation 39
• 4.8 Pixel Intensity Interval Analysis 40
• 4.8.1 Impact of Threshold Selection on Classification Performance 48
• 4.9 Model Performance Analysis Across Datasets 51
• 4.10 Visualization of Per-Class Results: Confusion Matrix and ROC Analysis 53
• 4.11 Performance Comparison Across Model Architectures 54
• 5 Conclusion and Future Directions 63
• 5.1 Concluding Remarks 63
• 5.2 Future Directions 64
• Bibliography 66
• 요약 72