When symptoms of cancer are observed, especially in conjunction with genetic
predisposition and/or lifestyle risk factors, a non-invasive exploration of the patient's body is required to confirm or refute the presence of a tumor. If there is evidence of a tumor there must be a formal diagnosis of the stage to which the cancer has progressed. 
 
The systemic analysis of the patients' biomedical data, which come from different biological levels, offer greater information for the medical decision process.
New biomedical technologies can allow the interpretation of the
origin of the illnesses, moving to an evidence-based medicine paradigm.

The recent increase in the complexity of the techniques for acquiring biomedical data as well as their innovative features, make it difficult for them to be incorporated in the clinician's practice.
Therefore, it would be convenient to develop automatic data processing and predictive models to objectively assist the experts to interpret the data in the decision workflow of diagnoses, prognoses, and treatments.

This Thesis focus its attention on the computer-assisted support of cancer diagnosis for clinical environments.
The aim of the study is to produce results with high accuracy in classification, transparency in relation to the clinical knowledge and capacity to generalize their performance to new samples subsequently obtained in different clinical centers.
The technical aspects covered in this Thesis includes the processing, modeling, feature extraction and combination of biomedical data;  the inference and evaluation of predictive models for biomedical problems; and
the integration of the  models into decision support systems for the clinical environments.
In order to focus our studies, two medical problems are tackled: Soft Tissue Tumor (STT) diagnosis and, Brain Tumor (BT) diagnosis. 

In the STT problem, high efficacy in the discrimination of the benign/malignant character of the tumors was achieved by Pattern Recognition (PR)-based classifiers on Magnetic Resonance (MR) Image findings. 
These classifiers can help radiologists the confirmation of the diagnoses of new patients, allow the study of suspicious cases, and aid in the education of new radiologists' expertise in tumors of this kind.
A new clinical decision-support system (CDSS) for STT has been designed and implemented, which is based on classifiers learned from multicenter datasets.
The generic distributed architecture designed for the STT problem has been the basis for posterior developments in the field, such as those adopted by the distributed CDSS of the HEALTHAGENTS project.

This Thesis provides several contributions to the BT medical problem.
A new approach that combines MR Spectra of different echo times has been proposed.
Significant differences in performance were found when Short TE, Long TE or both spectra combined were used as input.
In addition, a probabilistic mixture model and the E(xpectation)-M(aximisation)-based estimation for binned and truncated data with 
univariate mixture densities of means relative to a global shift 
have been proposed for Magnetic Resonance Spectroscopy (MRS) data characterization.
The new version of the mixture model keeps the biological information in the model and properly fits the MRS.
The discrimination of Brain Tumors based on the parametric space of the probabilistic mixture model is possible with high accuracies, and the combination of MR Spectra for classification can be performed by means of their parametric spaces of the models.
With respect to the evaluation of the PR-based models, it has been demonstrated that the prediction of in-vivo MRS cases
that are from a later date, from different hospitals, and with different instrumentation, but which are obtained under the same acquisition parameters may be possible by models inferred by multicenter datasets. 
Our results consolidate, with experiments on subsequently acquired datasets, the conclusions of previous studies on automatic brain tumor classification using MRS.
The classification results obtained by the PR-based models have been useful for the validation and review of cases with abnormal profiles in the multicenter datasets of the eTUMOUR project.
With respect to the classification of BT by ex-vivo data, a proof of principle was carried out using gene expression profiles to discriminate between glioblastoma and meningioma biopsies. 
The obtained gene-signatures are in accordance with the expected biological and pathological differences between these two kinds of tumors.

The conclusions and developments of this PhD thesis apply directly to the results of the eTUMOUR (Web accessible MR decision support system for brain tumour diagnosis and prognosis, incorporating in vivo and ex vivo genomic and metabolomic data, 2004-2009), and to the results of the HEALTHAGENTS project (Agent-based Distributed Decision Support System for Brain Tumour Diagnosis and Prognosis, 2006-2008), which are both European Union projects of the 6th Framework Programme.
Hence, based on scientific contributions studied in the Thesis for BT classification with in-vivo data, two practical solutions have been developed to integrate the PR engines in the clinical routine workflow.
The first one is a generic interface between the CDSS and the classification module developed for the eTUMOUR project which allows an on-line update of the available classifiers.
The second one is an agent-based distributed Decision-Support System, which is the basis for the European HEALTHAGENTS network to assist in brain tumour management.