Algorithms for Image Analysis and Combination of Pattern Classifiers with Application to Medical Diagnosis

Medical Informatics and the application of modern signal processing in the assistance of the diagnostic process in medical imaging is one of the more recent and active research areas today. This thesis addresses a variety of issues related to the general problem of medical image analysis, specifically in mammography, and presents a series of algorithms and design approaches for all the intermediate levels of a modern system for computer-aided diagnosis (CAD). The diagnostic problem is analyzed with a systematic approach, first defining the imaging characteristics and features that are relevant to probable pathology in mammo-grams. Next, these features are quantified and fused into new, integrated radio-logical systems that exhibit embedded digital signal processing, in order to improve the final result and minimize the radiological dose for the patient. In a higher level, special algorithms are designed for detecting and encoding these clinically interest-ing imaging features, in order to be used as input to advanced pattern classifiers and machine learning models. Finally, these approaches are extended in multi-classifier models under the scope of Game Theory and optimum collective deci-sion, in order to produce efficient solutions for combining classifiers with minimum computational costs for advanced diagnostic systems. The material covered in this thesis is related to a total of 18 published papers, 6 in scientific journals and 12 in international conferences.

💡 Research Summary

This thesis presents a comprehensive framework for developing Computer-Aided Diagnosis (CAD) systems in mammography, addressing the entire pipeline from image acquisition and feature definition to advanced pattern classification and classifier fusion.

The research begins by establishing a clinically-grounded foundation for diagnosis. Through collaboration with medical experts, a set of nine critical diagnostic features is identified, including tumor presence, microcalcifications, density, fat percentage, boundary vagueness, homogeneity, morphological shape type, patient age, and histological diagnosis. A dedicated mammographic database is constructed, with each case annotated using these features and verified by histology. Statistical analysis confirms the high diagnostic value of these features, particularly morphological shape, where stellate and micro-lobulated shapes show over 95% correlation with malignancy.

Concurrently, the work explores intelligent image acquisition hardware as part of the EU-funded I-ImaS project. It develops a system using a linear array of Monolithic Active Pixel Sensors (MAPS) capable of on-chip processing. This system employs a “scout scan” to perform real-time texture analysis and subsequently modulates the X-ray beam intensity during the main scan on a regional basis. The goal is to optimize diagnostic information in areas of interest while minimizing overall radiation dose. A texture-analysis model linked to exposure conditions is developed, identifying eight optimal textural features for feedback control.

The core of the diagnostic analysis involves extensive evaluation of pattern classifiers using the extracted features. Two primary data types are investigated: textural features (extracted at 20-pixel and 50-pixel box sizes) and morphological boundary features (analyzed using Radial Distance Signals, Discrete Fourier Transform, and Discrete Wavelet Transform). A wide range of classifiers—including Linear Discriminant Analysis (LDA), Least-Squares Minimum Distance (LSMD), K-Nearest Neighbors (K-NN), Radial Basis Function (RBF) and Multi-Layer Perceptron (MLP) neural networks, and Support Vector Machines (SVM)—are rigorously compared.

The results consistently demonstrate the superiority of non-linear classifiers. SVM and K-NN achieved the highest accuracy rates (up to 83.9% for texture-based classification and similar high performance for shape-based analysis), while linear classifiers (LDA, LSMD) performed significantly worse. This underscores the inherent non-linearity and complexity of the mammographic tumor discrimination task.

Finally, the thesis looks beyond individual classifier performance, proposing a novel framework for combining multiple classifiers under the principles of Game Theory. This approach seeks to achieve optimum collective decision-making by efficiently fusing the outputs of various classifiers, aiming to enhance final diagnostic accuracy and robustness while managing computational costs. In conclusion, the work provides a holistic set of algorithms and design philosophies for building efficient, accurate, and intelligent next-generation CAD systems for mammography, bridging the gap between image processing, machine learning, and clinical practice.