MY WORK
AMARA
X-Ray Quality Control
Breast Cancer Diagnosis
- Project Description
Deep learning has revolutionized artificial intelligence, with Radboudumc creating an algorithm for pulmonary nodule malignancy prediction that matches thoracic radiologist performance on low-dose chest CT.
To apply this algorithm in routine clinical practice, Radboudumc aims to enhance the current algorithm with additional CT data and will validate it against data from European lung cancer screening trials and five Dutch institutes.
The ultimate goal is that this AI tool will speed up the diagnosis of malignant nodules and reduce unnecessary nodule examinations.
Next to the development of this new algorithm, we aimed to further enhance its safety and trustworthiness with uncertainty estimation and out-of-distribution detection. Find a full description of my PhD project at DIAG Nijmegen. - Deep Learning Uncertainty Estimation
An important aspect for safe clinical implementation of a Deep Learning algorithm is the ability to gauge and communicate uncertainty. Uncertainty Estimation can help to identify situations where the algorithm has doubts about its predictions.
Leveraging this uncertainty estimation can optimize the clinical workflow by preventing the algorithm from making a lung cancer diagnosis in cases with high uncertainty and refer these the clinical experts for further evaluation.
This proactive integration of uncertainty estimation holds promise to minimize algorithm mistakes, thereby elevating the safety profile of a clinically adopted algorithm. - Out-of-Distribution Detection
Deep learning models may fail and suffer from reduced diagnostic performance when applied to unseen and abnormal data that varies from the training distribution. The reduced performance is even more evident on out-of-distribution (OOD) data that can be a result of changes in patient demographic, disease incidence or image acquisition and reconstruction. These changes are a phenomenon known as dataset shift.
Incorporating an OOD detection component into the Deep learning model can help in identifying dataset shifts and ensure safe and reliable clinical implementation.
- Background
The LRCB is the National expert center that performs the quality control of medical devices used in the Dutch Screening program for breast cancer and tuberculosis. The current quality control procedure of phantom images obtained with mammographic devices used in the Dutch breast cancer screening program is time consuming and observer dependent. During the quality control, physicists manually analyze phantom images for visible artefacts. With the use of deep learning algorithms, we aimed to classify phantom images with visible artefacts. - Experiments
In this project, we used several pretrained models to classify phantom images for a binary task (visible vs non-visible artefact) and a multi-class task (8 artefact classes vs non-visible). - Outcome
The trained models showed to be accuracte in the classification of artefacts for both the binary and multi-class classification task. These promising results demonstrated the potential deep learning has in the quality control of X-ray devices.
- Background
In the diagnosis of microcalcifications, linked to breast cancer, most patients undergo a biopsy since radiologists find it challenging to reliably discern between benign and malignant lesions. Most of these biopsies reveal benign results, causing unnecessary stress for the patients and increased diagnostic costs. With the use of machine learning algorithms, we aimed to develop a microcalcification malignancy risk estimation predictor. - Experiments
In this project, we extracted radiomics features from microcalcifications on low-energy contrast enhanced spectral mammography images to develop multiple machine learning models. The models were trained with and without including clinical parameters. - Outcome
The models showed an excellent distriminative performance in the classification of benign and malignant microcalcifications, which demonstrated its potential to improve clinical care and reduce unnecessary biopsies.
