Gharib, Sami
[UCL]
Bol, David
[UCL]
Lefebvre, Martin
[UCL]
In addition of being one of the most frequent types of cancer, colorectal cancer also represents one of the major causes of cancer deaths worldwide. Colorectal cancer can be diagnosed by detecting the presence of polyps inside the colon or the rectum, and due to its non-invasive properties, this is commonly done by using Wireless capsule endoscopy (WCE). In order to provide help to the doctors, WCE can be accompanied by computer-aided diagnosis (CAD). This allows saving time analyzing the images. On the other hand, deep learning has proven to be efficient when it comes to image recognition, and embedded inference has become popular as it answers concerns such as latency, security, and energy wasted on raw data transmission. In this work, we combine CAD and embedded inference by proposing a CNN based approach for embedded polyp detection inside a microcontroller. We started by designing a CNN for polyp classification, and we could achieve 4 times model compression to match microcontrollers memory constraints. This was done by reducing the model parameters to 8-bit fixed point numbers, at the cost of a 3% and 6% recall and accuracy deviation. The final model provides 92% recall and 90% accuracy on a dataset generated from the ETIS-Larib polyp database.
Bibliographic reference |
Gharib, Sami. Polyp classification for detection of colorectal cancer with capsule endoscopy. Ecole polytechnique de Louvain, Université catholique de Louvain, 2020. Prom. : Bol, David ; Lefebvre, Martin. |
Permanent URL |
http://hdl.handle.net/2078.1/thesis:26700 |