P1934 - Artificial Intelligence in the Identification of Colon Adenocarcinoma on Biopsies: The Dawn of a New Age

Mohammad Aloqaily, MD¹, Alexander Rabadi, MD², Nisreen Abu Shahin, MD³, Wafi Aloqaily, MD³, Safa Otoum, PHD⁴, Malek Aloqaili, MD³, Othmane Benlamlih, BSc⁴, Tala Ar'ar, MD⁵, Bahaa Abdelrahim, MD^5
1University of Maryland Medical Center, Baltimore, MD; ²McLaren Health Care, Michigan State University, Flint, MI; ³School of Medicine, The University of Jordan, Amman, 'Amman, Jordan; ⁴Zayed University, Dubai, Dubai, United Arab Emirates; ⁵University of Jordan, Amman, 'Amman, Jordan

Introduction: This study aimed to build artificial intelligence software using semi-supervised learning to establish a well-rounded convolutional neural network (CNN) to categorize histopathology slides from colon biopsies to either benign or malignant.

Methods: The methodology of this study was divided into two stages, with the first stage including the retrospective collection of histopathological slides from colon biopsies for patients either diagnosed with colon adenocarcinoma or who were found to have normal results, which resulted in the identification of a total of around 900 for each group.

The second stage involved the conversion of the microscopic slides into images using whole slide imaging and the generation of a language model using semi-supervised learning to establish a well-rounded convolutional neural network to detect the differences between the two groups after a meticulous preprocessing phase was initiated to ensure dataset homogenization. The preprocessing phase involved normalization techniques, extensive data augmentation, Selective color manipulations, and Various manipulations were also applied to introduce variability into the dataset, resulting in the development of a meticulously crafted CNN with the ability to accurately differentiate between colon adenocarcinoma and normal colon histology.

Results: The histogram analysis revealed significant variations in pixel intensities given that the CNN model underwent rigorous training leveraging the augmented dataset. The utilization of early stopping techniques prevented overfitting, leading to an observable progression in detection accuracy, increasing from approximately 58.7% and maxing at 87.9%. Concurrently, the metric loss showed a decline from 8.07 to 2.06, reflecting the model's learning curve.

Discussion: The average accuracy of 83.9% was noted. Nevertheless, the model accuracy increased dramatically from 58.7% and maxing at 87.9% with a concurrent decline from 8.07 to 2.06 in data loss, which signifies the growing potential of such a model with further training on larger data sets to optimize its accuracy and efficiency, which can provide additional insights into the model's performance and help to identify any potential weaknesses or biases and ensuring it is robust and adaptable to a wide range of real-world scenarios. The above results suggest that such models could be used as a valuable tool in medical applications, helping to improve the speed and accuracy of colon cancer diagnosis.



Disclosures:


Mohammad Aloqaily, MD¹, Alexander Rabadi, MD², Nisreen Abu Shahin, MD³, Wafi Aloqaily, MD³, Safa Otoum, PHD⁴, Malek Aloqaili, MD³, Othmane Benlamlih, BSc⁴, Tala Ar'ar, MD⁵, Bahaa Abdelrahim, MD⁵. P1934 - Artificial Intelligence in the Identification of Colon Adenocarcinoma on Biopsies: The Dawn of a New Age, ACG 2024 Annual Scientific Meeting Abstracts. Philadelphia, PA: American College of Gastroenterology.