Automated gastrointestinal tract classification via deep learning and the ensemble method
Colorectal cancer is a leading cause of death among the cancer family with a record of almost a million moralities in 2020 alone. While the treatment of colorectal cancer is very difficult, early diagnosis can help immensely with treatment, eliminating the risks, and recovery. In most cases early di...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | en en |
| Published: |
IEEE Computer Society
2021
|
| Subjects: | |
| Online Access: | https://umpir.ump.edu.my/id/eprint/33267/1/Automated%20gastrointestinal%20tract%20classification%20via%20deep%20learning_FULL.pdf https://umpir.ump.edu.my/id/eprint/33267/2/Automated%20gastrointestinal%20tract%20classification%20via%20deep%20learning.pdf https://umpir.ump.edu.my/id/eprint/33267/ https://doi.org/10.23919/ICCAS52745.2021.9649754 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Colorectal cancer is a leading cause of death among the cancer family with a record of almost a million moralities in 2020 alone. While the treatment of colorectal cancer is very difficult, early diagnosis can help immensely with treatment, eliminating the risks, and recovery. In most cases early diagnosis is possible by catching any of the precursors of the disease, many of which appear on the Gastrointestinal tract. The use of machine learning to automate the process of gastrointestinal tract examination could accelerate the process of diagnosis, and increase its efficiency. This study suggests the use of the stacking ensemble method with multiple pre-trained CNN models for an accurate classification of GI tract using the publicly available dataset Kvasir. The pre-trained models used in this study were ResNet50, MobileNetV2, and Xception, all of which were ensembled and trained on a subset of the data and tested on another to eliminate bias, and evaluates the model’s capacity for generalization. Overall, the model demonstrated impressive performance at 99.2% accuracy, 0.9977 AUC, and 99.29% F1-score, especially compared to the individual constituent models and other models discussed in the review section of the study. |
|---|