Enhancing diagnostic accuracy in neuroimaging through machine learning: advancements in statistical classification and mapping
Metadatos
Mostrar el registro completo del ítemAutor
Jiménez Mesa, CarmenEditorial
Universidad de Granada
Departamento
Universidad de Granada. Programa de Doctorado en Tecnologías de la Información y la ComunicaciónFecha
2023Fecha lectura
2023-10-27Referencia bibliográfica
Jiménez Mesa, Carmen. Enhancing diagnostic accuracy in neuroimaging through machine learning: advancements in statistical classification and mapping. Granada: Universidad de Granada, 2023. [https://hdl.handle.net/10481/85701]
Patrocinador
Tesis Univ. Granada.; Programa Predoctoral Formación de Profesorado Universitario (ref. FPU18/04902); Ministerio de Universidades (España); Programa de Proyectos de Investigación Precompetitivos para Jóvenes Investigadores del Plan Propio de la UGR (ref. PPJIB2021-14)Resumen
In recent years, the application of arti cial intelligence (AI) techniques in health
and medicine, including neuroimaging, has grown exponentially. Neuroimaging plays
a crucial role in studying the central nervous system and supporting clinical diagnosis
through non-invasive examination of the human brain. Computer-aided diagnosis
(CAD) systems have been developed to assist clinicians in this process by using pattern
recognition and prediction capabilities. These systems, created through multidisciplinary
collaboration, incorporate AI algorithms to improve diagnostic accuracy and
reduce clinician workload. However, these systems face certain challenges, such as the
lack of interpretability of AI models and the small sample sizes inherent in neuroimaging
studies, which complicate system learning and performance. Addressing these
challenges is crucial for establishing CAD systems as a standard for clinical diagnostic
support.
This thesis focuses on exploring various machine learning (ML) approaches to enhance
the accuracy and utility of CAD systems while ensuring optimal interpretability
for clinical analysis.
To enhance reliability, one approach involves tackling the curse of dimensionality by
reducing the number of features. The signi cance of feature selection and extraction
stages is emphasised in the development of an optimised multiclass classi cation
system. Additionally, validation methods implemented so far in neuroimaging studies
are questioned. The viability of an upper-bounded resubstitution as a validation method
is demonstrated through a non-parametric statistical inference framework, particularly
suitable in studies with small sample sizes. Moreover, the use of classical statistics in
neuroimaging, which usually rely on assumptions that are frequently violated, is also
questioned. To address this, a proposed data-driven approach for generating statistical
inference maps is tested in brain disorders such as Alzheimer’s Disease and Parkinson’s
Disease, and compared with a parametric approach.
Regarding interpretability, two systems are designed to provide easily interpretable
results for clinical experts. These systems place emphasis on the use of explainable AI
techniques. One system focuses on analysing sulcal morphology in individuals with
schizophrenia, while the other system proposes a method for detecting patterns in the
Clock-Drawing Test to assess cognitive impairment.
In summary, this thesis demonstrates the utility of ML techniques in neuroimaging
for brain mapping, feature detection, and classi cation. It explores the reliability and
interpretability of CAD systems, identi es potential improvements, and emphasises
the need for further research to develop techniques tailored to neuroimaging’s unique
conditions. This advancements will enable CAD systems to become a standard for
clinical diagnostic support, ultimately improving the quality of life for patients through
earlier and more accurate diagnoses