Abstract:
This study explores the effectiveness of Convolutional Neural Networks (CNNs) and
DenseNet models for binary and multi-class classification tasks using the Kvasir dataset,
which consists of gastrointestinal tract images. The research aims to perform a com
parative analysis of CNN and DenseNet models on both raw and refined datasets.
Initially, binary and multi-class classifications were conducted on the original dataset
to establish baseline performances for both models. Following this, K-Means Cluster
ing was applied for outlier detection to refine the dataset by removing anomalies. The
refined dataset was then used to re-evaluate the models’ performances on the same
classification tasks.
The results demonstrated that CNN outperformed DenseNet in binary classifica
tion tasks, achieving an accuracy of 91.1% on raw data and 91.0% on refined data,
while DenseNet’s accuracy dropped from 83.5% to 73.1% after refinement. For multi
class classification, DenseNet performed better on the raw dataset with an accuracy
of 84%, compared to CNN’s 74%. However, after data refinement, DenseNet’s perfor
mance decreased slightly to 79%, whereas CNN showed a minor improvement, achieving
78%. The data refinement process, involving outlier removal, did not significantly af
fect the overall performance of the CNN model but had a notable negative impact on
DenseNet, especially in binary classification tasks. These findings suggest that while
CNN is robust across both binary and multi-class tasks with or without data refine
ment, DenseNet is more sensitive to changes in the dataset. This study highlights the
importance of dataset refinement and its varying impact on different neural network
architectures in medical image classification.
