Abstract
The present thesis examines the usage of neural networks, in particular Convolutional
Neural Networks, for the detection of lung nodules in CT scans. The aim of this work
is the implementation of a nodule detection framework that allows the integration of
different datasets and models. For this purpose it initially provides a detailed overview
of the current state of research, thereby focusing on image preprocessing, data augmentation
and promising network architectures.
Employing the LNDb-dataset1 released in 2019 and an auspicious model architecture,
the training procedure of a lung nodule detection network is demonstrated. The model
is based on the architecture presented in 2 and uses a custom 3D-Single-Shot-Detector
as a detector head.
The model output is evaluated with the common object detection metrics precision and
recall, as well as with the FROC-score, which is a popular metric in medical applications.
We achieve a maximal recall of 0.68 at 42963 false positives per scan according
to the FROC-calculation procedure in the LNDb-Challenge3. The calculation of precision
and recall on the COCO IoU-range as well as the FROC-score for the typical set of
18
, 1
4 , 12
, 1, 2, 4, 8 FPs per scan yield a result of zero due to the low accuracy of box predictions
and the herefrom resulting lack of true positives. Thus the method presented
in this work is still not well enough trained in order to perform lung nodule detection.
However, more training time and the incorporation of additional datasets, image
processing procedures and training steps that are mentioned promise to lead to a significantly
better performance in the future.
Users
Please
log in to take part in the discussion (add own reviews or comments).