Evaluating the Performance of Densenet and Cocoholonet Models in PCOD Detection from Ultrasound Images
J. Prathibanandhi, G. S. Annie Grace Vimala · 2025
One common endocrine disorder in post-childbearing women is polycystic ovarian syndrome (PCOD). Some of the symptoms of the conditions which are associated with hormonal imbalances include: inability to conceive, irregular menstrual periods and hirsutism. Early recognition of the disease is mandatory if good control is to be achieved and to prevent the associated health risks. Herein, we describe a straightforward algorithm for diagnosing PCOD based on clinical and laboratory features. This method includes pre-processing and segmentation, then classification with the help of Cocoholonet. Rotation, flipping, zooming, and resizing are some of the image enhancements that are performed during the preprocessing stage. These techniques assist in bringing some measure or degree of homogenization in the inputs that are for analysis. Various segmentation strategies are applied in the usage of structures by fragmenting areas of interest. These approaches include the size of particles, contour, density, color, edge, texture, location, and watershed. This paper's study objective includes considering these important concerns when developing the deeper learning model Coco-HoloNet for precise PCOD recognition and classification in ultrasound pictures. These are multifaceted contributions that tackle the various issues with traditional methods to improve diagnostic accuracy. Convolution layers, density blocks, and appropriate pooling algorithms are all combined into CoCo-HoloNet, which makes it a very effective model for properly capturing and extracting crucial information from the input. In all four criteria, we measured—accuracy, recall, F1 score, and precision—COCOholonet consistently surpassed the competition when compared against Densenet 121 and Densenet201, among others. The results of our research show that COCOholonet continuously performs better than other models in terms of precision, accuracy, recall, and F1 score. Consequently, with an accuracy of about 98%.