This study uses deep-learning AI models to develop an unrestricted, computer-vision based approach that targets objective breast density in mastography screenings, focusing on convenience, reliability, and pertinence in underrepresented healthcare settings. The researchers implemented a custom-designed convolutional neural network (CD-CNN) and extreme learning machine (ELM) for image-based breast density classification purposes. The chosen dataset consisted of 10,372 full-field mammograms that were pre-classified by radiologists into one of the four Breast Imaging Reporting and Data System (BI-RADS) categories, from A through D. The 3 + ELM model, deemed to be the most promising, successfully concluded with a testing accuracy (percentage of correct classifications) of 95.4%, specificity (amount of predicted positives that are actually positive) of 98.0%, and sensitivity (correctly predicted positives) of 92.5%, indicating a strong overlap between the conclusions of the model and the researchers — the recorded weighted kappa was 90%, with the true kappa value falling between 82% and 98% according to a 95% confidence interval (CI). The overall final testing accuracy, specificity, and sensitivity are as follows: 0.739 (95% CI: 0.692–0.787), 0.811 (95% CI: 0.769–0.853), and 0.751 (95% CI: 0.704–0.798) respectively. The study not only concluded that pre-trained models demonstrated lower efficiency, but also that the structure of the 3 + ELM model was best suited for this dataset.