Titulo Estágio
Deep Learning for Automatic Segmentation and Classification of Adventitious Respiratory Sounds
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI-FCTUC
Enquadramento
Respiratory pathologies have increasing mortality and morbidity rates and are often associated with various comorbidities, such as cardiovascular diseases, metabolic syndrome, osteoporosis, mental illnesses and lung cancer.
Currently, there is no continuous follow-up of respiratory pathologies, leading to frequent episodes of exacerbation with hospitalization, which have a significant impact on the health and quality of life of patients, low productivity due to lost working days and high costs of outpatient treatment and hospitalization, imposing high public health burdens.
However, it is known that each episode of exacerbation is preceded by a phase of gradual increment that varies between several hours and several days before the peak of exacerbation. Thus, early detection of this trend towards peak exacerbation can prevent its occurrence and lead to a significantly milder clinical picture. As such, changing the current reactive paradigm in managing respiratory pathologies to a preventive, proactive and patient-centred one is essential. In this sense, systems for the early diagnosis of exacerbation are required.
In this context, a significant symptom is adventitious respiratory sounds (e.g., wheezes, crackles, etc.), often associated with respiratory disorders. In an environment of continuous monitoring, adventitious sounds should be detected automatically. However, this task becomes complex as lung sounds are easily masked by other respiratory sounds (e.g., deep breathing) and environmental sounds.
Objetivo
For several years, the Clinical Informatics Laboratory of CISUC has been developing systems for processing respiratory audio signals and image processing. This project is intended to develop algorithms for automatic segmentation and classification of adventitious respiratory sounds based on deep learning techniques.
The pursuit of the general objective set out is embodied in a set of objectives of a more specific nature, namely:
- Analysis and combination of public respiratory sound databases
- Study and application of deep learning techniques for the segmentation and automatic classification of adventitious respiratory sounds, either alone or together, in a controlled environment or with environmental acoustic disturbances.
Plano de Trabalhos - Semestre 1
1. Critical analysis of the state of the art related to the problem of segmentation and classification of adventitious respiratory sounds.
2. Critical analysis of the state of the art regarding the use of deep learning techniques in the segmentation and classification of sound in general (and respiratory sound in particular), e.g., CNNs, LSTMs, auto-encoders, etc.
3. Critical analysis of respiratory sound databases.
4. Initial segmentation and classification experiments
Plano de Trabalhos - Semestre 2
5. Research and development of deep learning approaches for the classification of respiratory sounds (including different pre-processing mechanisms, transfer learning mechanisms, application and adaptation of different deep neural network architectures, etc.).
6. Validation of the developed models and critical and comparative analysis of the results achieved.
7. Writing of thesis and scientific article.
Condições
- A shared deep learning server (with 8 GPUs) is available at our lab.
- Possibility of a research scholarship in the second semester.
Orientador
Rui Pedro Paiva, Diogo Pessoa, Paulo de Carvalho
ruipedro@dei.uc.pt 📩