The Effect of Batch Size, Shape, and Filter on CNN Architecture for the Classification of Cervical Intraepithelial Neoplasia Using Colposcopy Images
Zendi Zakaria Raga Permana, Agung Wahyu Setiawan · 2024
Colposcopy is a medical procedure that provides detailed visualization of the cervix and is non-invasive for the patient. Junior colposcopists only have sensitivity (55-64%) when performing colposcopic examination to determine the category of Cervical Intraepithelial Neoplasia (CIN) in patients. In addition, the accuracy of colposcopy examination result is essential for doctors’ decision-making in treating patients. Currently, numerous studies support decision-making in CIN using artificial intelligence. Existing research predominantly employs pre-trained transfer learning models, such as “imagenet”, which are trained on many images rather than specific CIN images. Therefore, this study proposes a classification approach that leverages pre-trained models on specific CIN images, utilizing a base Convolutional Neural Network (CNN) method. The parameters of the CNN are optimized by fine-tuning the filter kernel. The accuracy performance of the testing stage improved to 0.83, and the sensitivity increased to 0.84. The primary strategy is to combine the filter kernel parameters of the base CNN in the training stage, tuning the shape of filter kernel parameters to [128,64,32]. The first filter kernel is $5 \times 5$ and the second and third filter kernels are $3 \times 3$, resulting in a training accuracy performance of 0.83, validation accuracy of 0.61, and training time requirement of 66.28 minutes using Apple M1 Chip. In the future, a basic CNN will be developed using CIN data with green filters to create a model specifically used for the supervised identification of CIN characteristics.