Abstract:
The research in medical image classification using Deep Learning Techniques is a major
trend these days as it serves best in medical diagnosis. Deep Learning Models helps to
recognize the patterns with promising accuracy that task may be challenging for medical
practitioners due to time constraints and performing comparison with large datasets. As we
are testing the maximum limit of technology and going exponentially in the field. Similarly this
study was intended to reduce human effort by utilizing Deep Learning technologies in the field of
medical by classifying Gastrointestinal (GI) images from an open source Dataset KVASIR V2 with
all 8 classes. In most of the previous studies dataset was found as the main limitation as limited
number of samples and classes were used. As compare to previous research this thesis, does
image classification on increased number of classes with promising accuracy and identifies
the suitable parameters. Initial results were carried of 3 CNN based models out of which two
models EfficientNetV2B2 and VGG-16 were pretrained on ImageNet Dataset while the third model
was AlexNet. Evaluation metrics reported in this study include accuracy, precision, recall, f1-
measure in addition to categorical accuracies and Confusion matrix. Based on above metrics we
selected the best model EfficientNetV2B2 for further fine tuning. The study achieved promising
results in classification of GI Images based on pretrained model Efficient Net V2B2 with SGD
optimizer by achieving training accuracy of 97.03%, validation accuracy 94.03%, while
testing accuracy of 95.34%. Transfer learning technique was tested by utilizing above two pretrained
models as a foundation, transfer learning proved to be great for substantial reductions
in training time and processing resources beside AlexNet. Further utilizing Transfer learning
leaded to better training and improved performance in classification of Gastrointestinal (GI)
medical images. Lastly an application with GUI interface was built using Tkinter python library
to better interact with image classification process.
