Fracture detection and explainability in deep learning models
This thesis has been developed in the context of a collaboration project between health insurance company for work accidents Asepeyo and the Barcelona Supercomputing Center (BSC). The target of this project is the creation of a functional tool to support non-expert medical practitioners in the diagn...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/353720 |
| Acceso en línea: | https://hdl.handle.net/2117/353720 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine learning Diagnosis cnn explicabilitat gradcam lrp ciencies de la salut radiografia aprenentatge supervisat aprenentatge profund explainability healthcare radiography supervised learning deep learning resnet Aprenentatge automàtic Diagnòstic Àrees temàtiques de la UPC::Informàtica |
| Sumario: | This thesis has been developed in the context of a collaboration project between health insurance company for work accidents Asepeyo and the Barcelona Supercomputing Center (BSC). The target of this project is the creation of a functional tool to support non-expert medical practitioners in the diagnosis of radiology images. In this thesis, we focus the most common radiology diagnosed pathology in Spain: the wrist fracture. For this purpose, Asepeyo provided us with over 13,000 studies of real radiology data, as well as support from experts of the field for feedback. This thesis presents the problem and the analysis performed on it, as well as how we solve it using many AI methods both traditional and novel. Our main contributions include a complex preprocessing to deal with wrist images (as well as DICOM format images), training deep learning models in an environment with some degree of label uncertainty, obtaining models of reasonably high performance, and being able to unsupervisedly produce visual explanations on the radiologies that locate the fracture in the image - as this is a strong requirement for its potential use in practice. The results indicate that the model accuracy is enough to be deployed in a real world scenario, while the expainability outcomes illustrate the appropriate focus of the machine learning models on the fractures. Pending a few additional checks and bias validations, the outcome of this work could constitute a powerful and useful support tool for medical professionals non-experts in radiology. Lastly, we provide at the end a comprehensive list of the necessary requirements to be cleared before deploying the system into public healthcare. |
|---|