IntestiNet: Predicting Intestinal Abnormalities in Colon with Advanced CNN Techniques

Abstract

Background: In previous years, colon cancer seems to be the second leading cause of death worldwide, which needs preventive strategies to combat this disease. Thus, timely diagnosis with effective methods may improve the survival rate of patients.

Method: This article presents an advanced method as IntestiNet for the classification and prediction of colon cancer which leverages the abilities of CNN models in three major steps: (i) train and validate the CNN model with 54% training data (i.e. 3200 images) and 13% validation data (i.e. 800 images), (ii) parameter optimization using Adamax and custom callbacks to continue or halt training process, and (iii) test CNN models with 33% data (i.e. 2000 images). We herein created train_gen, test_gen and valid_gen using ImageDataGenerator class with image size of (200, 250) which further passed into six CNN models (EfficientNetB2, ResNet50, InceptionV3, VGG16, MobileNetv2, and EfficientNetB5) for the classification of four types of classes as ‘0: Normal’, ‘1: Ulcerative colitis’, ‘2: Polyp’, and ‘3: Esophagitis’.

Results: To validate the effectiveness of CNN models, an interactive approach using Keras applications for multiple hidden layer implementation has been conducted on a curated colon dataset. The results indicate how well the EfficientNetB5 model succeeds compared to other methods with an accuracy of 98.62%, precision of 98.62%, recall of 98.63%, and F1-score of 100%. Such outcomes demonstrate the possible benefit of EfficientNetB5 in enhancing colon cancer identification and diagnoses.

Conclusion: The proposed architecture in this article, demonstrates impressive propels in automatically identifying and classifying four types of colons. Comprehensive testing on a variety of metrics revealed EfficientNetB5's improved performance, which highlights the technology's potential to improve colon disease reconnaissance and prognosis.